Studies of non-diffusive heat conduction through spatially periodic and time-harmonic thermal ARCHIVES excitations MASSACHUSETTS INSTITUTE by APR 152015 Kimberlee Chiyoko Collins LIBRARIES Submitted to the Department of Mechanical Engineering in partial fulfillment of the requirements for the degree of Doctor of Philosophy in Mechanical Engineering at the MASSACHUSETTS INSTITUTE OF TECHNOLOGY February 2015 @ Massachusetts Institute of Technology 2015. All rights reserved. Signature redacted A uthor ........................ Department of Mechanical Engineering January 15, 2015 Signature redacted C ertified by.......................... Gang Chen Carl Richard Soderberg Professor of Power Engineering Thesis Supervisor Accepted by .............................. Signature redacted Cadu"Te ardt Chairman, Department Committee on Graduate Theses 2 Studies of non-diffusive heat conduction through spatially periodic and time-harmonic thermal excitations by Kimberlee Chiyoko Collins Submitted to the Department of Mechanical Engineering on January 15, 2015, in partial fulfillment of the requirements for the degree of Doctor of Philosophy in Mechanical Engineering Abstract Studies of non-diffusive heat conduction provide insight into the fundamentals of heat transport in condensed matter. The mean free paths (MFPs) of phonons that are most important for conducting heat are well represented by a material's thermal conductivity accumulation function. Determining thermal conductivity accumulation functions experimentally by studying conduction in non-diffusive regimes is a recent area of study called phonon MFP spectroscopy. In this thesis, we investigate nondiffusive transport both experimentally and theoretically to advance methods for determining thermal conductivity accumulation functions in materials. We explore both spatially periodic and time-harmonic thermal excitations as a means for probing the non-diffusive transport regime, where the Fourier heat diffusion law breaks down. Boltzmann transport equation calculations of one-dimensional (1D) spatially sinusoidal thermal excitations are performed for gray-medium and fully spectral cases. We compare our calculations to simplified transport models and demonstrate that a model based on integrating gray-medium solutions can reasonably model materials with a narrow range of dominant heat-carrying phonon MFPs. We also consider the inverse problem of determining thermal conductivity accumulation functions from experimental measurements of thermal-length-scale-dependent effective thermal conductivity. Based on experimental measurements of Si membranes of varying thickness, we reproduce the thermal conductivity accumulation function for bulk Si. To investigate materials with short phonon MFPs, we developed an experimental approach based on microfabricating 1D wire grid polarizers on the surface of a material under study. This work finds that the dominant thermal length scales in polycrystalline Bi 2 Te3 are smaller than 100 nm. We also determine that even small amounts of direct sample optical excitation, which occurs when light transmits through the grating and directly excites electron-hole pairs in the substrate, can appreciably influence the measured results, suggesting that an alternate approach that prevents all direct optical excitation is preferable. To study thermal length scales smaller than 100 nm without the need for microfabrication, we develop a method for extracting high frequency response information 3 from transient optical measurements. For a periodic heat flux input, the thermal penetration depth in a semi-infinite sample depends on the excitation frequency, with higher frequencies leading to shallower thermal penetration depths. Prior work using frequencies as high as 200 MHz observed apparent non-diffusive behavior. Our method allows for frequencies of at least 1 GHz, but we do not observe any deviation from the heat diffusion equation, suggesting that prior observations attributed to non-diffusive effects were likely the result of transport phenomena in the metal transducer. Thesis Supervisor: Gang Chen Title: Carl Richard Soderberg Professor of Power Engineering 4 Acknowledgments Many people contributed to the completion of this thesis in both tangible and intangible ways. I am deeply grateful to Prof. Gang Chen, who advised and supported my graduate studies. I am also grateful to the other members of my Ph.D. thesis committee, who spent many hours offering constructive feedback on my work: Prof. Millie Dresselhaus, Prof. Keith Nelson and Prof. Evelyn Wang. I was privileged to have the opportunity to learn from each of these extraordinary individuals. I also want to thank my colleagues who collaborated on the work presented in this thesis. In particular, I am grateful to Dr. Alex Maznev, whose close involvement was invaluable, and who I view as a scientific mentor. I would also like to acknowledge the contributions of Dr. John Cuffe, Vazrik Chiloyan, Lingping Zeng, Prof. Austin Minnich, Sam Huberman, Jeff Eliason, Prof. Keivan Esfarjani, Prof. Jivtesh Garg, Prof. Zhiting Tian, and Prof. Tony Feng. Beyond direct collaboration, I benefited from numerous discussions and interactions with my colleagues, especially Daniel Kraemer, Poetro Sambegoro, Dr. Ken McEnaney, Maria Luckyanova, Dr. Matthew Branham, Edi Hsu, Prof. Selcuk Yerci, Kara Manke, Sangyeop Lee, Dr. Sveta Boriskina, and Dr. Mayank Bulsara. Every member of the NanoEngineering Group at MIT impacted me and helped shape the direction of my graduate studies. I also benefited from those who came before me in the NanoEngineering Group, especially Prof. Aaron Schmidt, who mentored me at the start of my graduate career and taught me a great deal. The Spectroscopy Group within the Solid-State Solar-Thermal Energy Conversion (S 3TEC) Center also provided a forum for regular and fruitful discussion, and I am grateful to all those who participated. The resources offered at MIT also made this work possible. In particular, I would like to acknowledge the Microsystems Technology Laboratories and the NanoStructures Laboratory at MIT. I was aided especially by Kurt Broderick, Dr. Tim Savas, Jim Daley, Gary Riggott, Dr. Richard Hobbs, and Mark Mondol. Administrators and staff in the NanoEngineering Group, in the S 3TEC Center, and within MIT helped in 5 numerous ways both seen and unseen, and for that I am grateful to Ed Jacobson, Mai Hoang, Mary Ellen Sinkus, Juliette Pickering, Keke Xu, Leslie Regan, Read Schusky, Pierce Hayward, Mark Ralph, and Dick Fenner. I would like to thank my network of colleagues, friends and family who supported me throughout my graduate studies. I was privileged to be a part of MIT's MacGregor House as a Graduate Resident Tutor. I learned a great deal from the MacGregor community that I will carry with me for a long time. Many individuals and institutions mentored me and gave me opportunities that helped me attend MIT both as an undergraduate and as a graduate, especially Wes Masuda, Le Jardin Academy, 'lolani School, Dr. Gerry Luppino, Prof. George Ricker, Dr. Joel Villasenor, MIT Lincoln Laboratory, John Sultana, Prof. John Heywood, and Prof. Carol Livermore. I am deeply grateful for the support and friendship of Chris Celio and Dr. Ellen Chen. I owe thanks to all those I shared good times with, and to all those who helped me learn and grow. Finally above all else, I am forever grateful to my family, especially my parents and my siblings, whose unwavering love carries me through all of life's adventures. 6 Contents Introduction . . . . . . . . . . . . . . . . . . . . . 19 . . . . . . . . . . . . 21 . . . . . . . 23 Non-diffusive heat conduction 1.2 Thermal conductivity accumulation functions 1.3 Laser-based thermal property measurement techniques Frequency-domain thermoreflectance (FDTR) . . . . . . . . 24 1.3.2 Time-domain thermoreflectance (TDTR) . . . . . . . . . . . 25 1.3.3 Heat transfer model for TDTR and FDTR responses . . . . 27 1.3.4 Transient thermal grating (TTG) . . . . . . . . . . . . . . . 29 . . . 1.3.1 . . . . 1.1 Organization of the thesis 30 . . . . . . . . . . . . . . . . . . . . . 1.4 33 Boltzmann transport equation (BTE) overview . . . . . . . . . . . . 33 2.2 Modeling the TTG experimental geometry . . . . . . . . . . . . . . 35 2.2.1 Gray-medium BTE . . . . . . . . . . . . . . . . . . . . . . . 37 2.2.2 Spectrally-dependent BTE . . . . . . . . . . . . . . . . . . . 42 2.2.3 Frequency-integrated gray-medium model . . . . . . . . . . 46 2.2.4 Two-fluid model . . . . . . . . . . . . . . . . . . . . . . . . . 48 2.2.5 Contribution of long MFP phonons . . . . . . . . . . . 50 Thermal conductivity accumulation function reconstruction . . 51 2.3.1 Theoretical foundation . . . . . . . . . . . . . . . . . . 52 2.3.2 TTG experimental geometry . . . . . . . . . . . . . . . 2.3.3 TTG thin membrane experimental geometry . . . . . . 55 2.3.4 Experimentally deriving heat flux suppression functions 58 . . . . . . . . 2.3 . 2.1 . Modeling non-diffusive heat conduction . 2 19 . 1 7 . . . 53 2.4 . . . . . . . . . . . . . . . . . . . . . 63 3.1 Concept and design criteria for wire grid polarizer . . . . . . . . . . 64 3.2 M icrofabrication . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68 3.2.1 Dry etching Al 71 3.2.2 Polishing polycrystalline Bi 2 Te 3 samples . . . . . . . . . . . 73 3.2.3 Resulting one-dimensional grating structures . . . . . . . . . 75 . . grid polarizers . . . . . . . . . . . . . . . . . . . . . . . . . . . . Optical transmission results . . . . . . . . . . . . . . . . . . . . . . 76 3.4 TDTR results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78 3.4.1 Grating heat transfer model . . . . . . . . . . . . . . . . . . 80 3.4.2 Measurements varying pump laser diameter 81 3.4.3 Measurements varying angle between laser polarization and grat- . . . 3.3 . . . . . . . . . . ing transmission axis . . . . . . . . . . . . . . . . . . . . . . 81 Result summary . . . . . . . . . . . . . . . . . . . . . . . . . 84 3.5 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85 . Future directions . 3.4.4 4 59 Investigation of non-diffusive conduction with microfabricated wire . 3 Summary and future directions Frequency-domain representation of TDTR data, and applications for studying non-diffusive conduction 91 4.1 . . . . . . . . . . . . . . . . . . . . . . . . . . 94 4.1.1 FD T R . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94 4.1.2 Single-shot TDTR . . . . . . . . . . . . . . . . . . . . . . . . 95 4.1.3 TDTR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96 4.1.4 Frequency-domain representation of TDTR data . . . . . . . . 98 4.2 Theoretical foundation Experimental demonstration of fdTDTR . . . . . . . . . . . . . . . . 99 4.2.1 Data collection . . . . . . . . . . . . . . . . . . . . . . . . . . 100 4.2.2 Data stitching . . . . . . . . . . . . . . . . . . . . . . . . . . . 102 4.2.3 Fourier series representation . . . . . . . . . . . . . . . . . . . 103 4.2.4 Frequency upper limit 104 4.2.5 Direct comparison of fdTDTR and FDTR data . . . . . . . . . . . . . . . . . . . . . . 8 . . . . . . . . 107 4.3 5 Thermal model analysis of fdTDTR data . . . . . . . . . . . . . . . . 108 4.3.1 110 Sensitivity analysis . . . . . . . . . . . . . . . . . . . . . . . . 4.4 Comparison of fdTDTR data for various samples . . . . . . . . . . . 115 4.5 Summary and future directions . . . . . . . . . . . . . . . . . . . . . 119 Summary and Outlook 123 9 10 List of Figures 1-1 Conceptual illustration of how non-diffusive heat flux differs from the predictions of the Fourier heat diffusion equation. 1-2 21 Normalized thermal conductivity accumulation functions for various materials from first-principles calculations. 1-3 . . . . . . . . . . . . . . . . . . . . . . . . . . 22 Schematic diagrams of (a) the traditional frequency-domain thermore- flectance (FDTR) and (b) the broadband FDTR (BB-FDTR) systems in our lab. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1-4 Schematic diagram of the time-domain thermoreflectance (TDTR) system in our lab. 1-5 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31 Illustration of TTG crossed laser beam interference, which produces a spatially-sinusoidal temperature profile that decays in time. . . . . . . 2-2 26 Schematic diagrams of transient thermal grating (TTG) systems for (a) transmission and (b) reflection measurements. 2-1 25 36 Dimensionless gray-medium Boltzmann transport equation (BTE) thermal decays for a range of dimensionless length scales compared to diffusive limits and Fourier model best fits. 2-3 . . . . . . . . . . . . . . . . Dimensionless gray-medium BTE thermal decays compared to the ballistic transport lim it. . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-4 41 Normalized effective thermal diffusivities found by fitting gray-medium BTE decay curves with the heat diffusion equation. 2-5 40 . . . . . . . . . . 42 Set of material parameters for Si and PbSe at 300 K from density functional theory (DFT) which were used in our spectral BTE calculations. 11 44 2-6 TTG thermal decays for a range of grating periods calculated from a numerical solution to the spectral BTE using DFT input parameters for (a) Si and (b) PbSe at 300 K, compared to the diffusive limit. 2-7 . . 45 Effiective thermal conductivity for various TTG periods calculated from the spectral BTE for (a) Si and (b) PbSe at 300 K. Comparisons are shown for the gray-medium model using literature values of gray mean free path (MFP) and best fit gray MFP values. 2-8 . . . . . . 46 Effective thermal conductivity for various TTG periods calculated from the spectral BTE for (a) Si and (b) PbSe at 300 K. Comparisons are shown for the best fit gray-medium model and the frequency-integrated gray medium model. 2-9 . . . . . . . . . . . . . . . . . . . . . . . . . . . 47 Comparison of suppression functions from the two-fluid model and the frequency-integrated gray-medium model. . . . . . . . . . . . . . . . 48 2-10 TTG thermal decay calculated from a numerical solution to the spectral BTE using DFT input parameters for Si at 300 K, compared to the diffusive limit and the approximate solution provided by the two-fluid m odel. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49 2-11 Effective thermal conductivity for various TTG periods calculated from the spectral BTE for (a) Si and (b) PbSe at 300 K. Comparisons are shown for the best fit gray-medium model as well as the predicted results using both the gray-medium and the two-fluid heat flux suppression functions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50 2-12 Thermal conductivity accumulation functions for different grating periods calculated using DFT parameters for Si at 300 K and the two-fluid suppression function. . . . . . . . . .. . . . . . . . . . . . . . . . . . . 51 2-13 Percent contribution to effective thermal conductivity from phonons with mean free paths greater than half the TTG period. . . . . . . . 52 2-14 Reconstruction of thermal conductivity accumulation functions for (a) Si and (b) PbSe at 300 K using BTE calculated effective thermal conductivity values as "experimental" inputs. 12 . . . . . . . . . . . . . . . 54 2-15 Calculation of normalized effective thermal conductivity t(L) from reconstructed normalized thermal conductivity accumulation functions 4D(A) to verify agreement with the ,.(L) values used to reconstruct <b(A). 54 55 2-16 Illustration of a TTG measurement on a thin membrane. ....... 2-17 TTG measured effective thermal conductivity at 300 K for a range of Si membrane thicknesses. . . . . . . . . . . . . . . . . . . . . . . . . . 56 2-18 Fuchs-Sondheimer suppression function . . . . . . . . . . . . . . . . . 57 2-19 Reconstructed thermal conductivity accumulation function from the Fuchs-Sondheimer suppression function and experimentally measured effective thermal conductivities for thin Si membranes. . . . . . . . . 57 2-20 (a) Thermal conductivity per MFP and (b) thermal conductivity accumulation function for Si at 300 K from DFT calculations. . . . . . 59 2-21 Reconstructed heat flux suppression function for thin, diffusely scattering membranes. The exact solution given by the Fuchs-Sondheimer relationship is shown for comparison. 3-1 . . . . . . . . . . . . . . . . . . 60 Conceptual illustration of a one-dimensional metal grating acting as a linear polarizer to block pump and probe light from directly exciting a generic substrate. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65 3-2 Schematic illustrating the domain used for COMSOL calculations. . . 66 3-3 COMSOL calculations for transmittance and reflectance for 800 nm and 400 nm light for various grating geometries. . . . . . . . . . . . . 3-4 67 COMSOL calculations for transmittance and reflectance with a constraint on the relationship between L and d such that the gap between grating lines is held constant at L - d = 100 nm. 3-5 . . . . . . . . . . . 68 COMSOL calculations for transmittance and reflectance with a constraint on the relationship between L and d such that the filling fraction is kept constant with L = xd, where x is an integer. . . . . . . . . . . 3-6 69 COMSOL transmittance calculations varying the type of metal used for the one-dimensional grating. . . . . . . . . . . . . . . . . . . . . . 13 70 3-7 COMSOL transmittance and reflectance calculations for a Bi 2 Te 3 substrate patterned with a one-dimensional Al grating. . . . . . . . . . . 70 3-8 One-dimensional grating fabrication process flow. 71 3-9 Scanning electron microscope images of the one-dimensional grating . . . . . . . . . . . fabrication process. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72 3-10 Examples of post-etch corrosion (a) where corrosion products form clusters and (b) where corrosion results in removed sections of Al. . 74 3-11 Representative atomic force microscopy measurement of a polished polycrystalline Bi 2 Te 3 sample. . . . . . . . . . . . . . . . . . . . . . . 76 3-12 Examples of fabricated one-dimensional Al gratings. . . . . . . . . . . 77 3-13 Transmission measurement setup where either the 400 nm pump or the 800 nm probe is used . . . . . . . . . . . . . . . . . . . . . . . . . . . 78 3-14 Idealized linear polarizer transmitted intensity as a function of angle. 78 3-15 TDTR grating measurement setup. . . . . . . . . . . . . . . . . . . . 79 3-16 Measured substrate thermal conductivity as a function of pump laser diameter for (a) fused silica and (b) Si. . . . . . . . . . . . . . . . . . 82 3-17 Reflection results for various angles on (a) Si and (b) fused silica showing both amplitude and phase TDTR data as a function of delay time. 83 3-18 Thermoreflectance signal from bare Si. . . . . . . . . . . . . . . . . . 84 3-19 Reflection results for various angles on Si at (a) high transmission angles and (b) low transmission angles. . . . . . . . . . . . . . . . . . 85 3-20 Reflection results for angles around the low transmission angle for a one-dimensional Al grating on Si. . . . . . . . . . . . . . . . . . . . . 86 3-21 Concept for structures that prevent optical transmission into the substrate. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87 3-22 Using a crude model we extract substrate thermal conductivity values for substrates of Si and fused silica patterned with grating structures like that pictured in Fig. 3-21. . . . . . . . . . . . . . . . . . . . . . . 14 88 4-1 Illustration of thermal profiles in a semi-infinite solid that result from excitation by a sinusoidally periodic heat flux. 4-2 . . . . . . . . . . . . . Thermal penetration depth for semi-infinite solid slabs of Si, A12 0 92 3 and fused silica heated by a sinusoidal heat flux. . . . . . . . . . . . . 93 4-3 Typical laser heating profiles for (a) FDTR and (b) TDTR measurements. 94 4-4 TDTR experimental diagram. The additional static delay line com- bined with the movable delay stage allows for the collection of more than one full period of delay-time-domain data. 4-5 . . . . . . . . . . . . 101 Example of stitching TDTR data sets collected in two parts by introducing an additional fixed delay for measuring long delay time data (>6 ns)........ 4-6 ................................... 103 Room temperature TDTR data for a sample of A1 20 3 with an Al transducer layer, represented in (a) the frequency-domain and (b) the delaytim e-dom ain. 4-7 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105 A noise estimate is obtained by smoothing the raw delay-time-domain data, as shown in (a). Subtracting the smoothed curve from the raw data produces the noise data shown in (b). . . . . . . . . . . . . . . . 4-8 Comparison of the frequency response amplitude obtained from the raw data in Fig. 4-7(a) and the noise data in Fig. 4-7(b). 4-9 106 . . . . . . 107 Comparison of phase data from a frequency-domain representation of TDTR data (fdTDTR) and FDTR measurements on the same sample. 109 4-10 Thermal model best fits from simultaneously varying the A1 2 0 3 ther- mal conductivity, kAl 2 o 3 , and the Al-A1203 thermal interface conductance, G, are shown, assuming either a 10 nm or a 25 nm thick isothermal Al layer, dis.. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .111 4-11 fdTDTR phase data and model curves assuming isothermal Al layer . . . . . . . . . . . . . . . . 112 4-12 Thermal model sensitivity plots. . . . . . . . . . . . . . . . . . . . . . 113 thicknesses of 25 nm, 10 nm and 40 nm. 4-13 Contours of least squares fitting residuals given a range of kAl 2 o 3 and G values varied up to 10% about the best fit values. . . . . . . . . . . 15 114 4-14 fdTDTR data from Si, A1 2 0 3 , and fused silica substrates with a 110 nm thick Al transducer. Lines show thermal model assuming bulk properties and using di,, = 25 nm. . . . . . . . . . . . . . . . . . . . 117 4-15 fdTDTR data from Si, A1 2 0 3 , and fused silica substrates with a 60 nm thick Al transducer. Lines show thermal model assuming bulk properties and using dis, = 25 nm. . . . . . . . . . . . . . . . . . . . 118 4-16 fdTDTR data from Si, A1 2 0 3 , and fused silica substrates with a 160 nm Au transducer with a 5 nm Ti stiction layer. Lines show thermal model assuming bulk properties and using di 5 , = 80 nm. . . . . . . . 119 4-17 fdTDTR data from Si, A12 0 3 , and fused silica substrates with a 55 nm Au transducer with a 5 nm Ti stiction layer. Lines show thermal model assuming bulk properties and treating the entire Au layer as isothermal. 120 4-18 Comparison of early delay time TDTR amplitude signal shapes for different transducer layers on substrates of A1 2 0 3 including (a) 110 nm Al, (b) 60 nm Al, (c) 160 nm Au with 5 nm Ti, and (d) 55 nm Au with 5 nm T i. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121 4-19 fdTDTR data from a polycrystalline Bi 2 Te 3 sample with a 90 nm Al transducer layer. Model fits shown for diso = 25 nm and di,,, 16 = 10 nm. 122 List of Tables . . . . . . . . . . . . . 3.1 Fabricated grating polarizer extinction ratios. 3.2 COMSOL calculated extinction ratios for the fabricated grating polar- 3.3 79 izers in Table 3.1. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79 TDTR results on grating samples. . . . . . . . . . . . . . . . . . . . . 86 17 i8 Chapter 1 Introduction Heat conduction at small scales behaves differently from heat conduction at macroscales due to a lack of a well defined local thermal equilibrium. Phonons are quantized lattice vibrations that transport heat and sound in solids. Diffusive transport theory assumes a multitude of phonon scattering events which establish local thermal equilibrium, giving rise to a local temperature, and the temperature gradient governs the heat flux. When conduction length and/or time scales do not allow for many phonon scattering events, transport is non-diffusive, deviating from Fourier's heat diffusion law which over-predicts the heat flux in non-diffusive regimes [1, 2]. Studying heat conduction in non-diffusive regimes provides insight into the underlying nature of the phonons that carry heat in a material [3, 4], and can inform the design of materials with tailored thermal properties. The study of non-diffusive heat conduction to extract information about phonons is an emerging area of research called phonon mean free path spectroscopy [4, 5]. In this thesis, we explore both spatially periodic and time-harmonic thermal excitations as a means for probing the non-diffusive transport regime. 1.1 Non-diffusive heat conduction Phonons travel an average distance A, called the phonon mean free path (MFP), before being scattered. The MFP relates to the relaxation time 19 T through the group velocity v, where A VT. For length scales comparable to or smaller than phonon MFPs, or time scales comparable to or smaller than phonon relaxation times, energy transport deviates from the predictions of Fourier's law, which assumes local thermal equilibrium. Establishing local thermal equilibrium requires many local phonon scattering events. Heat transport is non-diffusive in lieu of sufficient scattering events to produce local thermal equilibrium. Fourier's law over-predicts the heat flux in non-diffusive transport regimes [1, 2], as illustrated conceptually in Fig. 1-1. For a material surrounded by parallel plate heat reservoirs at different temperatures, Fourier's law predicts a steady state heat flux of q = kAT/L, where L is the distance between the plates, AT is the indicated temperature difference, and k is the material's thermal conductivity. As L -+ 0, Fourier's law predicts a divergent heat flux, which is clearly unphysical for a finite AT. In reality, at some small enough L, the heat flux will reach a maximum value, as illustrated by the dashed line in Fig. 1-1. In this regime, the phonons travel between the hot and cold reservoirs totally ballistically without scattering, a phenomenon also called phonon radiative transport [6, 7]. In the ballistic limit, it is not possible to define a local temperature anywhere in the material. Between the ballistic limit and the diffusive limit, there exists an intermediate quasiballistic transport regime [3, 4], where some phonon scattering occurs, but not enough to define a local thermal equilibrium. Heat transport measurements that are performed in non-diffusive regimes are often interpreted using the Fourier heat diffusion equation, in spite of its lack of validity [5]. Measurement regimes near the onset of non-diffusive transport retain diffusive-like behavior [8], and can be modeled by the heat diffusion equation using a modified effective thermal conductivity, keff. The over-prediction of the heat flux by Fourier's law results in a lower value of experimentally determined effective thermal conductivity compared to that of the bulk material. The phonon Boltzmann transport equation (BTE) [9] is valid in both diffusive and non-diffusive transport regimes as long as wave effects are unimportant, but because the BTE is difficult to solve and requires a multitude of input parameters, the BTE is not typically used for 20 ballistic quasiballistic diffuse q| q = kAT/L L L L Figure 1-1: Conceptual illustration of how non-diffusive heat flux differs from the predictions of the Fourier heat diffusion equation (solid line). For 1D steady conduction between parallel thermal reservoirs separated by a distance L, the heat diffusion equation predicts a divergent heat flux q as L --+ 0, while in reality, the heat flux will reach a maximum at some small enough L where the phonons travel totally ballistically, as illustrated by the dashed line. directly fitting experimental data. We discuss solutions to the phonon BTE for simple experimental geometries in Chapter 2, as well as how BTE solutions can help link experimentally determined keff data to phonon spectral information. 1.2 Thermal conductivity accumulation functions Crystalline materials have distributions of phonon MFPs. One convenient representation of a phonon MFP distribution in a material is through the thermal conductivity accumulation function, kaccu, which sums the contributions of the thermal conductiv- ity per MFP, kaiff [10, 111, kaccu = kdiffdA, where 21 (1.1) P 3 L'ifQA(w dA (1.2) Here A is the phonon MFP, w is the radial frequency, p indexes the different phonon branches, C, is the mode specific heat, and v is phonon group velocity. The integration variable is A, so all properties are treated as functions of A rather than the more commonly used w dependence. Equations (1.1) and (1.2) assume isotropic dispersion relations and isotropic bulk MFPs [11. Thermal conductivity accumulation functions for several materials calculated from first-principles simulations are plotted in Fig. 1-2. From Fig. 1-2, we can see that half of the thermal conductivity of Si comes from phonons with MFPs > 600 nm, and that other materials like Bi have much shorter dominant heat carrying MFPs. 1.0 PbT a) Bi ~5 .B E rCoSb aA bS b :3 :3 0.6 0 0 S0 4 e Si o W0 0.20.0 T=300 K 1 10 100 1000 10000 phonon mean free path (nm) Figure 1-2: Normalized thermal conductivity accumulation functions for various materials from first-principles calculations from Ref. [12]. A material's thermal conductivity accumulation function shows which phonon MFPs dominate heat conduction in that material. The emerging area of phonon MFP spectroscopy [4, 5] aims to experimentally determine thermal conductivity accumulation functions of materials indirectly through measurements of the lengthscale-dependent effective thermal conductivity, keff(L), where L is the thermal length 22 scale in the measurement. Direct measurements of phonon lifetimes and dispersion relations are possible through inelastic neutron scattering [131, x-ray Raman scattering [14], or high frequency photoacoustics [15, 16]. While powerful as direct methods, neutron scattering and x-ray Raman methods require expensive equipment, usually in the form of national facilities, and are limited to single crystal samples. Photoacoustic methods have been used to measure frequencies up to 2 THz in superlattice structures [15], but generating high frequency acoustic waves in bulk samples remains a challenging limitation [17]. The cost and challenges associated with direct methods motivate the use of indirect methods to measure thermal conductivity accumulation functions. Measurements of length-scale-dependent effective thermal conductivity keff(L) data can be related to thermal conductivity accumulation functions kaccu(A), a process which will be discussed in Chapter 2. 1.3 Laser-based thermal property measurement techniques Laser-based, non-contact optical methods provide a convenient tool for studying heat transport. These techniques use a pulsed and/or modulated laser (pump beam) to heat the surface of a sample. The surface temperature variation leads to a change in optical properties, which is monitored by a probe laser beam. By comparing the measured response with model calculations, parameters of the sample such as the thermal conductivity and the thermal interface conductance between sample layers can be determined. In this section we introduce three notable measurement tech- niques that will be relevant for our investigations. Sections 1.3.2 and 1.3.4 discuss time-domain methods, and Section 1.3.1 discusses a frequency-domain method. Each of these has been used to study length-scale-dependent thermal conductivity, by varying either beam diameter [4], thermal penetration depth [18], or spatial interference 23 pattern period [19]. 1.3.1 Frequency-domain thermoreflectance (FDTR) The advent of using modulated light to excite a thermal response originates from the field of photoacoustics [20]. The use of thermoreflectance as a means for detecting a thermal response from a periodic heat input was pioneered by Rosencwaig and coworkers in 1985 [21], representing the first experimental demonstration of frequencydomain thermoreflectance (FDTR) [21, 22, 23, 18]. When a material undergoes a change in temperature, its index of refraction, and correspondingly its reflectivity, changes. Thus, observations of thermoreflectance can be related to a material's thermal response. A schematic of the FDTR system in our lab is shown in Fig. 1-3(a). A continuous wave (CW) pump laser beam is sinusoidally modulated by an electrooptic modulator (EOM). The pump produces time-harmonic heating on the sample, resulting in surface temperature oscillations at the modulation frequency, which are monitored by a CW probe beam. The amplitude and phase of the surface temperature response are measured as functions of the modulation frequency, making this a frequency-domain measurement. Phase measurements are oftentimes preferred due to their higher accuracy [18]. The modulation frequency in FDTR typically varies from kHz to -10 MHz [21, 22, 23]. Recently, an extension of the frequency range up to 200 MHz was reported [18, 24], using a method named broadband FDTR (BB-FDTR) by the developers. A schematic of the BB-FDTR system in our lab is shown in Fig. 1-3(b). By additionally modulating the reflected probe with a second EOM, BB-FDTR enables lock-in detection at the combined lower frequency of fo - fi, which allows for modulation frequencies up to fo = 200 MHz. Since the penetration depth of the temperature oscillations becomes smaller at higher frequencies, such an extension is beneficial for studying thermal transport at fine spatial scales. 24 (a) laser iCW 488 nm pump EOM, fo CW lasrobe BS sample obj. n V4 PBS r_ band pass detector filter (b) CW laser LJ 488 nm pump EOM, fo CW laser 532 nm probe BS sample obj. X/4 PBS EOM, f1 band pass detector filter Figure 1-3: Schematic diagrams of (a) the traditional frequency-domain thermore- flectance (FDTR) and (b) the broadband FDTR (BB-FDTR) systems in our lab, which were based off the design of Ref. [18]. 1.3.2 Time-domain thermoreflectance (TDTR) Time-domain thermoreflectance (TDTR) is a closely related technique to FDTR. The first TDTR experiment was reported by Paddock and Eesley in 1986 [25], and notable improvements to the technique were made by Capinski and Maris [26 and by Cahill and coworkers [271. In TDTR, the pump beam comes from a femtosecond laser operating at a high repetition rate (typically -80 MHz) and is additionally modulated by an EOM, as illustrated in Fig. 1-4, which depicts the TDTR setup in our lab. Unlike FDTR, the heating in TDTR is not time-harmonic but is comprised of many frequency components. The probe beam is derived from the same laser, and the probe pulses are delayed with respect to the pump pulses by a mechanical delay line. Our TDTR setup uses a probe beam that is concentric with the pump beam. To 25 prevent pump light from entering the detector, we frequency-double the pump beam using a second harmonic generator (SHG) optic. To mitigate overlap alignment errors as the delay stage is swept, the probe is expanded prior to entering the delay line, and the focused probe beam diameter on the sample is typically much smaller than the pump beam diameter. The thermoreflectance response is measured by a lock-in amplifier with the pump modulation frequency serving as a reference. The TDTR signals, i.e. the in-phase and out-of-phase (quadrature) outputs of the lock-in amplifier as functions of the delay time, do not, in general, reflect the actual time-domain dynamics of the surface temperature of the sample. However, just as in FDTR, the response can be compared to model calculations to determine the properties of the sample such as the thermal conductivity of the substrate and the thermal boundary resistance between the substrate and the metal film typically used to facilitate both laser-induced heating and thermoreflectance measurements [28, 29]. fo red SHG filter EOM, fo 4x compress 44x pump 400 nm probe 800 nm A/2 sample LLJ U 1Ox dichroic BS obj. expnd exan f~ 111 moable molasae delay stage aser, fs blue detector filter Figure 1-4: Schematic diagram of the TDTR system in our lab [30, 29]. 26 k+ k2 1.3.3 Heat transfer model for TDTR and FDTR responses Typically TDTR data is interpreted by solving the heat diffusion equation in the frequency domain to find the temperature response to time-harmonic heating, just as in FDTR measurements. In order to model a TDTR signal, many frequency responses are added together [28, 29]. Modeling TDTR and FDTR responses with the thermal diffusion equation is well documented in the literature [28, 29, 31]. We follow the methodology of Ref. [29] to model heat transport using a thermal quadrupole [32, 33] solution to the heat diffusion equation for a multilayer, semi-infinite solid, which accounts for both radial and anisotropic conduction. The heat diffusion equation in cylindrical coordinates is C OT ot (1.3) r Or r Or &z2 where T is temperature, k is thermal conductivity, C is volumetric heat capacity, r is the radial in-plane direction, z is the cross-plane direction, and t is time. The time-harmonic nature of the heat flux excitation and the radial symmetry of the problem enable simplifications of the heat diffusion equation with Fourier and Hankel transforms respectively. Applying Fourier and Hankel transforms to Eq. (1.3) leads to iwCT = -krS 2 T + kz (z2' (1.4) where w is the radial frequency Fourier transform variable, s is the spatial Hankel transform variable, and T is the temperature in Fourier-Hankel space. Rearranging, 2 =-mT, where m 2 = (kS 2 (1.5) + iwC)/kz. The in-plane direction (r-direction) is assumed to be isotropic, while anisotropy in the cross-plane direction (z-direction) is accounted for. Equation (1.5) is a second-order ordinary differential equation, and fits nicely into the thermal quadrupoles framework [32, 33]. 27 The thermal quadrupole approach readily adapts to multilayer structures by the relating heat fluxes and temperatures on the top and bottom surfaces of a multilayer stack through matrixes of material properties. For a material layer, cosh(md) -sinh(md) M, =,,m -kn sinh(md) (1.6) cosh(md) where d is the layer thickness. For an interface layer between materials, the property matrix becomes I G-1 0 1 Ain = , (1.7) where G is the thermal interface conductance, which is the inverse of the interface resistance. A relationship between the Fourier-Hankel-domain temperatures T and heat fluxes q on the top and bottom surfaces of a structure with n layers is given by 1A Tbottom jI qo (1.8) ~qoton qbottom = M7 1M Ai~...Kf 1 [ B T 0P C D qOP , where the top layer corresponds to n = 1. For TDTR, the top surface is thermally excited and probed, and the bottom layer is thick enough to be semi-infinite. For a semi-infinite substrate, no heat flux exists through the bottom surface, so the top surface temperature, Ttp relates to the top surface heat flux, qtop, in the FourierHankel-domain by TOP= -D ~ C(1.9) For time-harmonic heating with a Gaussian-shaped pump laser beam, the top surface heat flux input is qtOP - 2AO X -2r2 28 exp (ioot) , (1.10) where A, is the absorbed pump power, wo is the pump 1/e 2 radiusi, and w, is the excitation frequency. Taking Fourier and Hankel transforms of Eq. (1.10) leads to Z Ao qtop = -- exp -s, ((8 (1.11) -w,). Thus from Eq. (1.9), the surface temperature in Fourier-Hankel space T',,p is Z TtOP = -D Ao C 27 /s2 xp 8 - o ).(1.12) The response is measured by a coaxial probe beam that also has a Gaussian intensity distribution with a 1/e 2 radius of wi. Taking an inverse Hankel transform and weighting by the probe intensity distribution leads to a frequency-domain surface temperature of [28, 29 A top(w) = s 6(W - wo) 2.7r -D) -s(_82W C (1.13) 0exp 9)ds. 8 The frequency response to a time-harmonic heat flux input is given by h(w) = D08-s2( N(W) 0 o s +2(W (_ exp 8 ) T(w)/(w). Thus the measured frequency response h(w) is proportional to [28, 29] (1.14) 72)ds. For practical purposes, constants in front of h, such as the absorbed laser power, are unimportant because they either cancel out, as in the case of analyzing the phase of h, or they are not experimentally determinable so the data is treated without absolute units, as in the case of analyzing the amplitude of h. 1.3.4 Transient thermal grating (TTG) Transient thermal grating (TTG) experiments [34, 35, 36] generate a spatially sinusoidal optical grating by interfering two short pulsed coherent pump beams. The angle of interference between the beams 0, and the pump wavelength A, determine 'The 1/e 2 radius defines the radius at which the intensity is reduced to 1/e2 intensity at the center of the beam. 29 = 0.135 of the peak the spatial grating period L, where L = A/(2 sin(O/2)) in a medium of air. Absorption of the optical grating induces a periodic temperature profile that causes a spatially periodic change in the refractive index, which depends on temperature. The resulting diffraction grating is examined with a probe beam. The temperature profile relaxes in time as heat flows from the peaks to the nulls of the spatial thermal grating, causing the intensity of the diffracted probe to vary in time. The rate of the transient grating relaxation T, depends on the material thermal diffusivity a, and the grating 9 wavevector q = 27r/L, where T = aq2 Figure 1-5 illustrates TTG experimental setups for both transmission and reflection style measurements [37]. A phase mask is used to split the pump beam into 1 diffraction orders, which are subsequently focused on the sample, producing a spatial interference pattern. The angle of interference 0 can be adjusted by varying the period of the phase mask. The diffracted probe is mixed with a low intensity reference beam for phase-controlled heterodyne detection [38], which improves signal-to-noise. The relative phase of the probe and reference beams are adjusted by rotating a glass slide in the probe path. The detector signal as a function of time is monitored with a high resolution oscilloscope. 1.4 Organization of the thesis In this thesis, we explore various strategies for measuring non-diffusive transport and how such measurements can be related to thermal conductivity accumulation functions, which provide valuable information about the phonon MFPs which are most important for transporting heat. In Chapter 2 we discuss modeling non-diffusive transport using the Boltzmann transport equation and reconstructing thermal conductivity accumulation functions from experimental measurements in non-diffusive transport regimes. Our BTE calculations focus on TTG experimental geometries, which are especially amenable to theoretical analysis. We explore different simplified models for approximating the full BTE result, and show that integrating gray-medium solutions can reasonably reproduce solutions for materials with narrow thermal con- 30 532 nm probe = 515 nm pump - (a) ND filter sample detector phase phase mask probe + reference adjust (b) 532 nm probe detector - 515 nm pump ND filter sample 0.phase mask phase adjust Figure 1-5: Schematic diagrams of TTG systems for (a) transmission and (b) reflection measurements based off of Ref. [371. ductivity accumulation functions. We also use measured TTG data on Si membranes of varying thicknesses to accurately reproduce the thermal conductivity accumulation function of bulk Si. Chapter 3 presents our investigation of a method for measuring non-diffusive transport over 100 nm length scales using microfabricated wire grid polarizers. We design gratings that meet a set of criteria including light blocking, fabrication constraints, and likely requirements for observing non-diffusive behavior. We fabricate gratings on substrates of Si, fused silica and polycrystalline Bi2 Te 3. Transmission measurements on the transparent fused silica samples confirm the polarizing characteristics of our fabricated gratings. TDTR measurements reveal clear non-diffusive behavior in Si and bulk behavior in fused silica, as expected, although we identify that some small direct optical excitation of the Si substrate adds to experimental uncertainty. For polycrystalline Bi 2 Te3 , we do not observe any deviation from bulk thermal con- 31 ductivity, suggesting that the heat carrying phonons in polycrystalline Bi 2 Te 3 have MFPs of less than 100 nm. To study even smaller length scales without the need for microfabrication, we develop a method for extracting high excitation frequency information from our TDTR measurements, which we discuss in Chapter 4. We show that FDTR and TDTR provide identical information, and develop a method for transforming TDTR data into the frequency response data measured by FDTR. The high time resolution inherent in TDTR systems enables the extraction of high frequency response information, at least as high as 1 GHz, which is a regime currently inaccessible to FDTR methods. Our measurements to date have not observed behavior that deviates from the heat diffusion equation, suggesting that earlier reports of non-diffusive transport at high frequencies [18] may have resulted from transport effects in the metal transducer layer, which may obscure observations of non-diffusive transport in the substrate under study. 32 Chapter 2 Modeling non-diffusive heat conduction In non-diffusive transport regimes, the assumptions in the Fourier heat diffusion equation break down, as discussed in Section 1.1. A more appropriate model that applies in both diffusive and non-diffusive regimes is the Boltzmann transport equation (BTE) [9, 2]. To gain intuition about behavior in non-diffusive transport regimes, we examine BTE solutions for the simple transient thermal grating (TTG) experimental geometry introduced in Section 1.3.4. We go on to explore methods for connecting experimentally measured thermal length scale information to thermal conductivity accumulations functions, which were introduced in Section 1.2. Reliable methods to measure thermal conductivity accumulation functions are a key challenge for the field of phonon mean free path spectroscopy. 2.1 Boltzmann transport equation (BTE) overview A general form of the BTE is given by [9, 2] Of Of Of = where f scatt f+v-f+a- (2.1) is the distribution function, v is velocity, a is acceleration, and t is time. For 33 phonon heat conduction, the acceleration term drops out because there are no forces accelerating phonons. The scattering term can be simplified under the relaxation time approximation, which is valid for temperatures comparable to or greater than the Debye temperature [39]. Under the relaxation time approximation, the scattering term becomes Of fo - f where T is the relaxation time. pation function, f0 , (exp(hw/kBT) - 1 (2.2) T Ot)scatt For bosons (like phonons) the equilibrium occu- follows the Bose-Einstein distribution function, f, = fBE = )-i, where h is the reduced Planck constant, w is angular fre- quency, kB is Boltzmann's constant, and T is temperature. For heat transport along one dimension (ID) where boundary scattering is unimportant, the BTE becomes 2o- Of + i f = -fo 0--, f(23 +11 at OX T (2.3) where x is the spatial dimension and i = cos(6) is the directional cosine. When using the BTE to describe phonon dynamics, it is convenient to perform a change of variables, defining the phonon distribution function g as the phonon energy density per unit frequency interval per unit solid angle, g = hwD(w)f/4Tr, where D(w) is the phonon density of states. Thus, the phonon BTE (also referred to as the equation of phonon radiative transport and the Boltzmann-Peierls equation) is given by [40, 9, 2] Bg Og - + pt at OX where v is the phonon group velocity and g -g = T T , is the phonon relaxation time. (2.4) The equilibrium distribution function follows the Bose-Einstein distribution. A first-order Taylor expansion can be used to further simplify the equilibrium distribution go, leading to a proportional relationship with temperature [41] 34 1 go(T) = -- hwD(w)fBE eg0 o(T) 47r 1 - 1C 47r (2-5) (T - To), where C, is the differential, frequency-dependent specific heat (also called the mode specific heat). Such a simplification is reasonable as long as the temperature deviations from the equilibrium temperature, To, are small. In addition, conservation of energy requires [9, 42, 41 0jwrn I f J (2.6) didw. dpdw = In the gray-medium approximation, where a material contains a single phonon relaxation time and no parameters depend on frequency, the integrals over W cancel, leading to an even simpler relationship between g and go, g0 - 1j 2 _ (2.7) gdp. Together, Eqs. (2.4) and (2.6) are used to solve for the equilibrium distribution, which relates to temperature, for all space and time given a set of initial and boundary conditions. 2.2 Modeling the TTG experimental geometry One notable experimental method for measuring thermal properties is the transient thermal grating (TTG) technique [34, 35, 36], which we introduced in Section 1.3.4. The TTG technique is especially amenable to theoretical analysis because of its inherent spacial symmetry. As shown in Fig. 2-1(a), crossed laser-beam interference produces a sinusoidal pattern with a period L determined by the laser wavelength A and the angle of interference 0, where L = A/(2 sin(0/2)) in air. Absorbed optical energy leads to a sinusoidal temperature excitation that decays in time following the short pulse laser excitation, as illustrated in Fig. 2-1(b). For opaque samples, the excitation grating exists near the surface, while for transparent samples, the grating penetrates throughout the depth of the sample. 35 Our analysis treats the one-dimensional case of a TTG measurement in a bulk material not influenced by boundary scattering, where the depth of the thermal grating is much greater than the grating period. Shortly after pulsed laser excitation, the temperature profile is given by T(x,t = 0) = Ta.cos(qx), (2.8) where T is the temperature deviation from T, the average background temperature, Tmax is the peak temperature deviation, t is time, x is the spatial variable, and q = 27r/L is the spatial wave vector. The solution to the heat diffusion equation yields an exponential decay of the form T(x, t) = Tma cos(qx)e-a 2 , (2.9) where a is the material's thermal diffusivity. Lt (b) (a) Figure 2-1: (a) TTG crossed laser beam interference produces a sinusoidal interference pattern that induces (b) a spatially-sinusoidal temperature profile that decays in time. Non-diffusive phonon transport has been observed when L is within the range of the mean free paths (MFPs) of the heat carrying phonons in a material [19]. At small grating periods, Johnson et al. observed deviations from the bulk, diffusive thermal conductivity of Si membranes. Our theoretical treatment differs from the experimental geometry of Ref. [19], where heat transport was influenced by the boundary scattering in a thin membrane. Rather, our approach is appropriate for TTG measurements in bulk materials, where the depth of the thermal grating is 36 much greater than the grating period [34]. In order to extend this approach to thin membranes or strongly absorbing materials in which heat is dissipated into the depth of a sample, a multidimensional BTE would need to be considered. We discuss a boundary scattering problem for thin membranes in Section 2.3.3. In non-diffusive transport regimes, where the assumptions of the heat diffusion equation break down, the transport physics are better described by the BTE. We begin by considering solutions to the gray-medium BTE under the relaxation time approximation in Section 2.2.1, before moving on to discuss spectral solutions in Section 2.2.2. 2.2.1 Gray-medium BTE The simplest approach to solving the BTE is to assume a gray-medium, where only one phonon MFP exists in a material, and both v and T are constants. For the gray- medium BTE, we derive a dimensionless, semi-analytical solution for the ID TTG relaxation, and show the limiting behaviors of the decay. As described in Section 2.1, under the relaxation time approximation, the 1D phonon BTE takes the form of Eq. (2.4) and obeys Eq. (2.7). By Eq. (2.5), for small temperature excursions relative to the background temperature T,, the equilibrium distribution g, is proportional to temperature g 0 (Xt) 47r hwD(w)fBE(T) 47r CwT. (2.10) For mathematical convenience, the distribution function g and the temperature T are defined as deviations from the average background values, which correspond to thermal equilibrium at the background temperature. We seek a spatially-periodic solution for g by assuming that g and T are of the form p(t, p)exp(iqx) and T(t)exp(iqx). Now Eq. (2.4) takes the form of a first-order ordinary differential equation for g, OtT + go= ,(2.11) 37 with the solution j(t, pt) 1 - e-('-')jo(t')dt' + Ae-ft, (2.12) (1 + iqprVT)/T and A= j(t = 0,where [t) = j,(O). Although I the thermal grating is one-dimensional, we account for phonons traveling in all directions using the directional cosine, t = cos(O). Applying Eq. (2.7) leads to o(t')e(t'-t)/rsinc(qv(t' - t))dt', jo(t) = jo()sinc(qvt)e-/ + -j T0 where sinc(x) = sin(x)/x. We can further generalize Eq. (2.13) (2.13) by substituting nondimensional variables, t (=-, T 27rA )T y~v L L 0 (t) - go(0) _ T(t) ,(.4 (0)( producing T(() sinc(r()eC + T((')sirnc(j((' - ())e('-C)d(', (2.15) which is a nondimensional solution of the gray-medium, one-dimensional phonon BTE for a sinusoidal temperature profile. Here A is the phonon MFP, which relates to relaxation time and group velocity through A = vr. 9 is the phonon Knudsen number, a ratio of MFP to characteristic length. Here the characteristic length is given by the spatial wave vector of the grating, q = 27/L. Equation (2.15) is a Volterra integral equation of the second kind, which can be solved using standard numerical techniques [43]. An alternate approach to deriving a non dimensional, analytical solution to the ID gray-medium BTE for a TTG temperature profile assumes a spatially-periodic solution for g and utilizes a Fourier transform in time. Starting with the gray-medium phonon BTE with a spatially-periodic instantaneous source, 38 = +p + o (2.16) A6(t)eiqx, assuming a periodic spatial solution of the form, exp(iqx), and taking a Fourier transform in time leads to where = = z 9 -- (217 ( 2.17 A, T ) -ivy + iqv p= f j(t, p)eivtdt. Integrating over p to satisfy energy conservation, as shown in Eq. (2.7), gives AT z 9 qA (tan-' (j-0) - (2.18) 1 which is an analytical solution for the equilibrium distribution function in the frequency domain. To obtain a time domain solution, an inverse Fourier transform can be performed numerically, - A-T A = pdv. (2.19) 27r exp(-ivt) -oc qA (tan-' ( -1 A)) Writing Eq. (2.19) in the nondimensional variables of Eq. (2.14) leads to T(C) where 2 = = 127 d', _) 7O( tan- (2.20) -1 vT. Equations (2.15) and (2.20) produce identical decay curves, shown by the solid lines in Fig 2-2. A range of different 77 values are used to compute T(() decays. Small rj values show more diffusive behavior, while larger n values produce highly non-exponential decays. Decays that oscillate about T = 0 (the n = 5 curve for example), indicate that the sinusoidal temperature profile relaxation illustrated in Fig. 2-1(b) overshoots the T, equilibrium position before settling. The solution to the heat diffusion equation, given by Eq. non-dimensional variables as 39 (2.9), can be rewritten in terms of our T(() = e-O, where i (2.21) = aq 2T. The diffusion equation predicts an exponential decay described by 0. In the diffusive limit, abulk = vA/3, and / 3 ul =l rj2 /3. For comparison, the diffusive limit curves are also plotted in Fig. 2-2. The diffusive limit curves yield faster decays than the corresponding BTE curves, indicating that the thermal transport at small length scales slows down compared to Fourier law predictions [9, 1, 2, 8, 411. -- gray BTE -- diffusive lim. 0..5 05 =5 0 2 4 6 8 10 Figure 2-2: Dimensionless gray-medium BTE thermal decays for a range of dimensionless length scales (solid lines) compared to diffusive limits (dashed lines) and Fourier model best fits (dot-dash lines). At large r7 values, the decay becomes strongly non-exponential and acquires an oscillatory character. The thermal decay for very large r/ values approaches the ballistic limit, as shown in Fig. 2-3. In the ballistic limit, T -4 oo, and Eq. (2.15) reduces to T = sinc(r](). We can understand this oscillatory behavior by considering the case of purely ballistic transport, which would be equivalent to having non-interacting particles moving with a constant velocity v, and having an initial density distribution cos(qx). For a subset of particles whose velocity makes an angle 0 with the x direction, the particle density will oscillate as cos(qx - qvpt). Integrating over all angles yields a sine function in time, identical to the ballistic limit derived from the BTE. 40 71= q = 100 10 1 66 -gray 0.5 - S0 BTE -ballistic lim. -- 0 2.5 0.5 5 0 - t1T 1 = t1T Figure 2-3: Dimensionless gray-medium BTE thermal decays compared to the ballistic transport limit. We can also gain insight by fitting the gray-medium BTE decays using the heat diffusion solution to find best fit values of /, or correspondingly, the "effective" thermal diffusivity. In experiments, analyzing non-diffusive data with the Fourier heat diffusion equation is common practice [37, 19]. Example best fit Fourier curves are shown in Fig. 2-2. This produces a set of effective values of 0, which are normalized and plotted in Fig. 2-4. In the limit of small r, which corresponds to large grating periods, transport is in the diffusive regime, with Oeff /#buk 3 3 / eff /712 - Ceef/ Ceul = 1, and Fourier fits are good. At progressively larger values of I, the transport transitions to the ballistic regime, with BTE curves displaying highly non-exponential behavior which cannot be captured by the fitted Fourier curves, and the effective diffusivity approaches zero. This result does not mean that ballistic phonons do not carry heat; they simply transfer much less heat than diffusion theory predicts. The curve in Fig. 2-4 is universal due to its dimensionless form, and the graymedium assumption. If a material behaved like a gray-medium, we could predict the aef values that would be measured in a TTG experiment by scaling the horizontal axis with the appropriate phonon MFP. In general, however, materials do not behave as gray-mediums, so it is more accurate to consider a solution to the BTE that accounts for different phonon modes. 41 - - 1 0.8-50.6.0 ao0.4 0.20 2 10 -1 10 0 1 10 10 7=27rA/L 2 10 3 10 Figure 2-4: Set of effective thermal diffusivities aeff normalized to the bulk thermal diffusivity abulk, which were found by fitting the gray-medium BTE decay curves for different values of q with the exponential solution from the heat diffusion equation. 2.2.2 Spectrally-dependent BTE Since real materials support a range of phonon MFPs, a solution to the spectrallydependent BTE will provide a more accurate representation of the real thermal relaxation as compared to the simpler gray-medium approximation. Moreover, a solution that incorporates a realistic phonon density of states and set of phonon relaxation times would be an improvement over semi-empirical models, like the Callaway [441 or Holland [45] models. We utilize density of state and relaxation time data for all six phonon branches calculated from first principles density functional theory (DFT) in our spectrally-dependent BTE solution [46, 47]. Figure 2-5 shows the DFT parameters utilized in our calculations for Si and PbSb respectively at 300 K. The range of values plotted arise from different directions in the Brillouin zone. Our calculation assumes an isotropic material, so we interpolate the DFT parameters to find the mean value for a given phonon frequency and branch, and verified that our averaged properties produced literature values for the thermal conductivities and volumetric specific heats. Si and PbSe are interesting case materials to consider due to their different thermal conductivity accumulation functions. Si has a thermal conductivity accumulation function spanning a wide range of MFPs, 42 from tens of nanometers to tens of microns [46], while PbSe has a much more narrow distribution [47]. In the 1D spectrally-dependent treatment, the distribution function depends on four variables, g g(x, t, p, w). As in the gray-medium case, we assume a spatial dependence of the phonon distribution and temperature of the form exp(iqx), producing 0+ qt ot - + 'tqp/g = g -gj (2.22) 2.2 T T The equilibrium phonon distribution function is found from Eq. (2.6). To simplify the analysis of Eq. (2.6), a small temperature rise is assumed, such that - .1 o(T) ~~-Cwi, (2.23) 4fr leading to a temperature variation of [41] T 0' j -dw fo j f_1 T (2.24) dpdw. The summation over phonon branches is implied in the integrations over W. Our numerical solution of Eqs. (2.22)-(2.24) uses an explicit finite difference scheme. The granularity used in our calculations for the variables in the finite difference solution included at least 32 bins for p and 100 bins for w, with dt < 2 ps. Convergence was verified by systematically increasing granularity. We further verified our spectral BTE code by inputting gray-medium parameters and achieving identical results to the gray-medium model discussed previously. As an additional confirmation, we implemented a solution that did not assume a spatial dependence on g or T, which produced the same results, but at a much higher computational cost. Figure 2-6 shows temperature decay curves for Si at 300 K for a range of TTG periods, and the corresponding diffusive limit decays. As with the gray-medium model, the thermal grating decays more slowly than the diffusive model predicts, even for grating periods as large as L = 20 pm. The decay retains an exponential behavior even at grating periods as small as L = 1 pm, in contrast to the gray-medium 43 10 8 TA1 TA2 *LA - LO 10000. 0 4 8000- 00 o 0 8e 00o a)F E - TA2 * LA - LO TOl 10 -TO1 *** * 6000- 10 00 1 X10 o TA1 0. -TO2 0.61 0W E en 0 .4 3 - 5 o) (rad/s) t 10 10-1 0 5 o) (rad/s) X 1013 TOl TA1 TA2 LA 9 2NA -1 2000 ^0 04C C T021I -1 4000 10 8- - 12000 LO 0 .2 0 10 x 1013 5 (0 (rad/s) 10 x 1013 (a) 4500 - TA1 . TA2' * LA * TOl 4000 3500 3000 200 %* 10~10 o TA1 TA2 OLA - T01. 02 N 0* 15T02 1 x 10TA2 10 9 .*0 0.- 0 .8 cc 0 C? E Ci) LA 0 .4 P 101 :!!F T02 0 .6 TA1 10 > LO TOl 2040 0 .2 25- . -13LI (t 0 (rad/s) X 1013 2 o (rad/s) 4 x 1013 0 0 2 o (rad/s) 4 13 x 10 (b) Figure 2-5: Set of DFT material parameters for (a) Si and (b) PbSe at 300 K, including phonon group velocity v, relaxation time T, and mode specific heat C", for longitudinal acoustic (LA), longitudinal optic (LO), transverse acoustic (TA), transverse optic (TO) phonon branches as a function of phonon frequency w. Different data for the same value of w arise from different Brillouin zone directions. DFT data was provided courtesy of Keivan Esfarjani and Zhiting Tian [46, 47]. 44 solution shown in Fig. 2-2, which exhibits non-exponential behavior even for small values of rq = 27rA/L. L = 2 rm L =0.2 Rm 1~ L = 20 sm Si I -spec. BTE -- -diffusive lim. ( 0.5 0 o 0.1 0.2 0 t (ns) 5 t (ns) 10 0 250 t (ns) 500 (a) L = 100 nm L=40nm L=10nm 1 -spec. BTE -- -diffusive lim. PbSe CD 0.5 0 0.01 t (ns) 0.02 0 0.1 t (ns) 0.2 0 0.5 t (ns) 1 (b) Figure 2-6: TTG thermal decays for a range of grating periods calculated from a numerical solution to the spectral BTE using DFT input parameters for (a) Si and (b) PbSe at 300 K, compared to the diffusive limit. Again proceeding in accordance with typical experimental methodologies, we fit the BTE decays using the heat diffusion equation to find effective values of the thermal diffusivity for each grating period. The resulting effective diffusivities are normalized and plotted in Fig 2-7(a) for Si and Fig.2-7(b) for PbSe. Now that we have calculated a spectral BTE solution, we can compare it to the simpler gray-medium BTE, for which we derived the universal solution plotted in Fig. 2-4. Normally, MFP estimates are obtained from the experimental values of thermal diffusivity using the expression given by the gray-medium BTE, abuIk = vA/3, and assuming the Debye model in which v is the branch-average acoustic velocity [2]. This approach yields MFP values of ~40 nm for Si [2, 48] and -2 nm for PbSe [47, 49], but these produce effective diffusivity curves shifted towards much lower grating periods than our spectral BTE calculations, as shown in Fig. 2-7. It has been suggested that the gray-medium BTE can be made to work better for Si by using a larger MFP value 45 0.8 m 0 0.6 >0.4 -, >0.4 ,| 0.2 0.2 , 010 ----- r 10 10-9 10 < 10~ 1 L (m) Ie Si, 300 K S . ,,*. .. *...5.-------3 / gray, A =1 4mPbSe, 300 K -- 1pc 0.8 ....... +spec, BTE o.--gray,A=2nm .. gray, A = 6.5 nm 10 10 110 ~910~ 10~1 10 - 10-5 10~4 10 L(m) (a) (b) Figure 2-7: Effiective thermal conductivity for various TTG periods calculated from the spectral BTE for (a) Si and (b) PbSe at 300 K (solid markers). Comparisons are shown for the gray-medium model using literature values of gray MFP (dashed lines) and best fit gray MFP values (dotted lines). [50, 48, 18]. We find that the best fit to the spectral BTE results for Si is achieved with a MFP as large as 1 pm, and even then the fit is quite poor, as can be seen from the dotted line in Fig. 2-7(a). For PbSe, a MFP of 6.5 nm yields a somewhat better fit to the spectral BTE results, as shown in Fig. 2-7(b). In the next section, we discuss approximate models that are able to more accurately match our spectral BTE results. 2.2.3 Frequency-integrated gray-medium model A single-MFP gray-medium model is not able to accurately describe the (L) that we calculate from the full spectral BTE. We can improve on the single-MFP model by assuming that phonons of frequency w, for a given phonon branch, contribute to the thermal conductivity according to the gray-medium model with MFP A(w). The effective thermal conductivity is found by summing over the phonon spectrum as follows: keff K where the function Sgray kbf1 kwuhr 1 1 f" 3kbutko Jo Sgray ovAdw, (2.25) describes how the contributions of phonons are reduced 46 compared to the predictions of the diffusive heat diffusion equation. Sgray = aeff/abuk is shown in Fig. 2-4. This approach might be reasonable if phonons of different frequencies did not interact such that phonons at each frequency, for a given phonon branch, obeyed the gray-medium BTE. At room temperature, phonon scattering is dominated by phononphonon interactions, in which case Eq. (2.25) lacks a solid foundation. Nevertheless, one can hope that it will yield an improvement over the single-MFP gray-medium model, and indeed we observe that this "frequency-integrated gray-medium" approach does yield better results, as shown by the dash-dot lines in Fig. 2-8. In fact, for PbSe the dependence of the effective diffusivity on grating period is reasonably reproduced over a wide range of grating periods. The thermal conductivity accumulation function for PbSe is more akin to a graymedium than the accumulation function for Si, which spans a much broader range of phonon MFPs. To describe the behavior in Si, we require a model that accounts for the interactions of different phonon modes. 1 0.8 - - spec. BTE --- gray, A= 1 1 -' ..- 0.8 _+0.6 10.6 -9 -5 0.4 I 0.4 - .- 0.2----10 1 10 1 L (m) 1 Si, 300 K 10 i PbSe, 300 K *0m -- freq-int gray-med. +0 . spec. BTE gray, A = 6.5 nm -freq.-nt. gray-med. . 5 10 10 1 1 1104 L (m) 10 (b) (a) Figure 2-8: Effective thermal conductivity for various TTG periods calculated from the spectral BTE for (a) Si and (b) PbSe at 300 K (solid markers). Comparisons are shown for the best fit gray-medium model (dotted lines) and the frequency-integrated gray medium model (dash-dot lines). 47 2.2.4 Two-fluid model Maznev et al. derived an approximate solution to the spectral BTE in the ID TTG geometry by focusing on the onset of non-diffusive transport, where the TTG period is much larger than the MFPs of high-frequency phonons that are responsible for most of the specific heat [8]. Accordingly, those high-frequency phonons are assumed to obey the diffusion model, whereas the low-frequency phonons are analyzed with the BTE. Within this two-fluid approach, it was found that the TTG decay remains exponential, as in 2.9, with the thermal conductivity modified by a suppression function as follows: keff kbulk 1 _ 3 kbu1k J fWmax (2.26) with Stwo-fluid where q = 27rA/L. 4 ( Figure 2-9 compares -- , tan' Sto fluid to (2.27) Spay from our frequency- integrated gray-medium model. 1 -sgray C/) c 0.8 two-fluid 0.6 0 0.6 U) 1 -1 0 Uc -2 10 0 100 10 102 103 fl = 27cA/L Figure 2-9: Comparison of suppression functions from the two-fluid model and the frequency-integrated gray-medium model. A comparison of our calculated spectral BTE decays for Si at 300 K and the 48 exponential TTG decays predicted by Ref. [8] shows good agreement down to L =1 pm, where the spectral BTE yields a nearly exponential decay, and remains reasonable even at L = 0.2 pm, as shown by the dotted lines in Fig. 2-10. In Ref. [8] it was suggested that for Si at 300 K, the approximate solution would be expected to work for L > 1 pm. We see that in fact it works quite well for a much wider range of TTG periods. L 0.2 [m BTE -spec. -- diffusive lim. - two-fluid L=1 m . 0 1 1L 0.1 t [ns] 0.20 =2im 1 t [ns] L =20 2 m 0.5 0 5 t [ns] 100 250 t [ns] 500 Figure 2-10: TTG thermal decay calculated from a numerical solution to the spectral BTE using DFT input parameters for Si at 300 K (solid lines), compared to the diffusive limit (dashed lines) and the approximate solution provided by the two-fluid model (dotted lines). The two-fluid model proposed by Ref. [8] closely predicts the dependence of effective diffusivity on TTG period for Si at 300 K, as shown by the dashed line in Fig. 2-11(a). The agreement is quite good in Si for L > 1 pm, and is reasonable in PbSe for L > 200 nm. The assumptions in the derivation of the two-fluid model are only valid during the onset of non-diffusive transport, where the grating period is large compared to the MFPs of high frequency phonons that are primarily responsible for specific heat. These high frequency phonons are modeled as a thermal reservoir that obeys the heat diffusion equation, while low frequency phonons, which are mainly responsible for thermal conductivity, are modeled with the BTE. Thus, the two-fluid model is only expected to be valid for large grating periods, which agrees with the 49 findings in Fig. 2-11. spec. BTE 0.8 e PbSe, 300 K ------ gray, A= 1 stm ---- freq.-int. gray-med. --- two-fluid 0.8,4/ 0.6 / _0.6 0.4 0.4 0.2 0.2 i + spec. BTE 10~ //- freq.-int. gray-med. 10 10 810 10 L (m) gray, A = 6.5 nm /... Si3 K two-fluid '-- Si, 300 K 10~5 10 -4 10 10-9 10 10~ 10L (m) 10-5 104 10- (b) (a) Figure 2-11: Effective thermal conductivity for various TTG periods calculated from the spectral BTE for (a) Si and (b) PbSe at 300 K (solid markers). Comparisons are shown for the best fit gray-medium model (dotted lines) as well as the predicted results using both the gray-medium (dot-dash lines) and the two-fluid (dashed lines) heat flux suppression functions. 2.2.5 Contribution of long MFP phonons Experimental observations of large deviations from the heat diffusion equation do not necessarily indicate that long MFP phonons are contributing significantly to the observed heat flux reduction. In fact, for typical TTG periods (L > 1 pim), most of the reduced heat flux can be attributed to phonons with A < L/2. As a first approximation, we define long MFP phonons as those with A > L/2 and consider the contributions of these phonons to reductions in thermal conductivity. The thermal conductivity accumulation function, kaccu is defined as kaccu(A) = SovA 3 dA dA, (2.28) where S is a function describing how the phonon modes are reduced relative to the predictions of the diffusive heat diffusion equation as a function of experimental length scale. In Section 2.2.4, we demonstrated that the two-fluid model is in good agreement with the full spectral BTE calculations for Si at 300 K, so for simplicity we use 50 S = Stwo-fluid, given by Eq. (2.27). Figure 2-12 shows kaccu(A) for several different TTG periods using DFT input parameters for Si at 300 K. Each curve asymptotes at the effective thermal conductivity plotted in Fig. 2-11(a) for a given grating period. 150 .6 bulk L= 100 pm 100- L= 10pm E a 5Q .L=1 pm CO) L=0.1 pm 10 10 10 10 10 MFP,A ([im) 10 10 Figure 2-12: Thermal conductivity accumulation functions for different grating periods calculated using DFT parameters for Si at 300 K and the two-fluid suppression function (see Eq. (2.28)). Dotted lines indicate A = L/2. We examine the contribution to kaccu provided by phonons with A > L/2. Figure 2-13 shows the percentage contribution of phonons with A > L/2 to kaccu as a function of grating period. The contribution of these long MFP phonons is small until the TTG period is smaller than 100 nm, even though significant reductions in keff compared to kbulk 2.3 are observed at grating periods as large as 50 pfm, as shown in Fig. 2-11(a). Thermal conductivity accumulation function reconstruction In Sections 2.2.4 and 2.2.3 we discussed the forward problem of predicting effective thermal conductivity values for different TTG periods from known material parameters and a model of heat flux suppression. More useful is the inverse problem of predicting thermal conductivity accumulation functions from measurements of lengthscale dependent effective thermal conductivity and a model of heat flux suppression. Solving the inverse problem allows experimental measurements to be predictive of 51 ; 0.8 0 0 ~0.6- o o Z"0.4A 0.20 LI- *-" 2--3o 10-3 2 -- 10-2 1 0 - 102 10-1 100 101 TTG period, L ([m) 103 Figure 2-13: Percent contribution to keff from phonons with A > L/2. material properties, namely thermal conductivity accumulation functions, which have been shown to be important for analyzing thermal transport in bulk materials and nanostructures [10, 11]. 2.3.1 Theoretical foundation Minnich [51] suggested a way to reconstruct the thermal conductivity accumulation function, 4D(A) = kaccu(A)/kbulk, using non-diffusive measurements of normalized effective thermal conductivity, i(L) = keff(L)/kbulk, and an appropriate suppression function, S, as follows: (L) = OS( )(A)dA= JOJ 0 D(A)dA. (2.29) #(A) through <D(A) = d77 dA Here FD(A) is related to the thermal conductivity per MFP f #(A')dA'. S is a function of a dimensionless length scale (oftentimes the phonon Knudsen number) q = A/L, where L is the thermal length scale in the experiment and A is the phonon MFP. A similar equation appears in the analysis of thermal conductivity size effects in nanostructures, with the nanostructure dimension defining L [11]. 52 Even though such a reconstruction is an ill-posed problem, given the limited number of 4(A). K measurements, progress can be made if certain constraints are imposed on Minnich [51] showed that if o(A) is a smooth function that monotonically increases from 0 to 1, convex optimization [52, 53] can be used to reasonably estimate D(A). 2.3.2 TTG experimental geometry Following Minnich's approach [51], we reconstruct the thermal conductivity accumulation functions for Si and PbSe at 300 K. For K(L), we use our full spectral BTE solutions plotted in Fig. 2-11. For S(r), we test both the two-fluid model, Stw,_flu2d (given by Eq. (2.27)) and the frequency-integrated gray-medium model, Sray (shown in Fig. 2-4). For both suppression function models, the dimensionless length scale is given by y = 27rA/L, where L is the TTG period. The resulting reconstructions are shown in Figs. 2-14(a) and 2-14(b). For the reconstruction, we used a smoothing factor of 1, and set the length of P to 200 elements. We found that modifying the smoothing factor or the length of (D by a factor of 2 had little effect on the resulting reconstruction. The reference D(A) distributions are calculated from the same DFT dispersion and relaxation time data [46, 47] that we used for our spectral BTE calculations. We can verify the reconstruction by using the resulting 4(A) to calculate K(L) from Eq. (2.29). The resulting calculations, compared to spectral BTE solutions for K(L), are shown in Fig. 2-15. The reasonable agreement provides confidence in the convergence of the convex optimization algorithm. As expected from the results plotted in Fig. 2-11, using Stofluiid produces a more accurate MFP distribution reconstruction for Si and using Sgray produces a more accurate reconstruction for PbSe. The approximations in the two-fluid approach of Ref. [8] lead to a more accurate result for low-frequency, long MFP phonons, and indeed, we observe that the two-fluid model well reproduces the long MFP thermal conductivity accumulation function for both Si and PbSe. Our frequency-integrated graymedium approach is more appropriate for materials that approximate gray-mediums 53 .------- - 1 Si, 300 K PbSe, 300 K 1a 0.8 0.8 0 0 0.6 .9 0.4 00 0.6 0 0 0.4 - 0.2 )-reference refe rence a o 4)fr oom S,, o ( fr oom Stwo-fluid 0.2 - n 1-9 1- 1 0 -6 -7 Do 0 0-4 0-5 o rom S gray a 4 from 10 -3 A (m) 10__11 10 -10 10-9 10 A [m] (a) 10 10 1-7 (b) Figure 2-14: Reconstruction of thermal conductivity accumulation ftnctions for (a) Si and (b) PbSe at 300 K using BTE calculated effective thermal conductivity values as "experimental" inputs. Reconstructions using a gray-medium (open circles) or a twofluid (open squares) heat flux suppression function are shown. Thermal conductivity accumulation functions calculated from DFT are also shown for reference (solid lines). 1 0. . 1' * spec. BTE o from S o3 from S 0. 8 S 0 0. 6- 0. 6 0. 4 - 0. 4- 0 0. 2 10 10- 10^ 10-5 10-4 + spec. BTE fromSgray 4) reconst. O 0. 2 S Si, 300 K 0 10~10 10 PbSe, 300 K D reconst. 40 reconst. 3 from Stflud 4P reconst. * 10 -3 L (m) '- 10 10 10 -9 10 -8 10 -7 10 -6 10 -5 10 -4 10 -3 L (m) (a) (b) Figure 2-15: Calculation of normalized effective thermal conductivity i(L) from reconstructed normalized thermal conductivity accumulation functions <D(A) to verify agreement with the ri(L) values used to reconstruct <D(A) (closed diamonds). Verification is shown for <b(A) determined from Sgray (open circles) and from StOaflicd (open squares) for both (a) Si and (b) PbSe. 54 with step-like thermal conductivity accumulation functions, and hence works better for PbSe than for Si. 2.3.3 TTG thin membrane experimental geometry A convenient geometry for TTG measurements is that of a thin membrane, where the optical penetration depth of the pump and probe lasers is longer than the membrane thickness, allowing for a transmission-style measurement. Such a scheme is illustrated in Fig. 2-16. pump beams Figure 2-16: Illustration of a TTG measurement on a thin membrane [54]. Measurements on membranes have the advantage of a well defined thermal length scale, determined by the membrane thickness, that can be varied over a wide range. Recent TTG measurements on Si membranes ranging in thickness from 15 nm to 1.5 pm were reported [54]. The data spans a wide range of effective thermal conductivity, making it especially amenable to reconstructing the thermal conductivity accumulation function. The experimentally measured normalized effective thermal conductivity values as a function of membrane thickness are plotted in Fig. 2-17. This data was collected at a large TTG period, such that the dominant thermal length scale was the membrane thickness, not the grating period. Measurements with different grating periods produced the same results and are included in the plotted error bars. 55 1 0.8 =3 *0. 6 0.4- 0.2 00-9 10 10 -8 -7 10 d(m) 10 -6 10 -5 Figure 2-17: TTG measured effective thermal conductivity at 300 K for a range of Si membrane thicknesses [54]. A membrane geometry is also amenable for theoretical analysis. The analytical Fuchs-Sondheimer suppression function was derived from the BTE for electron transport in thin films [55, 56], and has been adapted for phonon transport [19, 54]. The Fuchs-Sondheimer suppression function is given by I -I)e-x'7dx, S(-) (2.30) 1- q+ j where r = A/d and d is the membrane thickness. The suppression function is shown in Fig. 2-18. Using SFS in Eq. (2.29), we reconstruct the bulk thermal conductivity accumulation function in Si, and achieve good agreement with the results from DFT, as shown in Fig. 2-19. Since there is some variation in the measured data, as shown by the error bars in Fig. 2-17, we perform a series of reconstructions using all the possible combinations of maximum and minimum K data error bars to determine the resulting restriction variation, which is plotted as error bars in Fig. 2-19. 56 1 0.8 Cl) UCD) 0.6 - -3 - 0 0.4 0.2 0 1ic -4 10 10 10 10 10 10 10 10 1 = Aid Figure 2-18: Fuchs-Sondheimer suppression function, SFS, where d is the membrane thickness and A is phonon MFP. 1 0.8- --I-- & 12 -refere nce D o D frorr SFS 0.60.4- 0.2[ 10 If .90.- -8 -7 -6 -5 -4 -3 -2 10 10 10 10 10 A(m) 10 10 10 Figure 2-19: Reconstructed thermal conductivity accumulation function <((A) (open circles) from the Fuchs-Sondheimer suppression function SFS and experimentally measured K for thin Si membranes (see Fig. 2-17). The thermal conductivity accumulation function from DFT calculations [46] is shown for comparison (solid line). Error bars in the reconstructed 4D come from reconstructions based on error bars in the measured K data. 57 2.3.4 Experimentally deriving heat flux suppression functions If a heat flux suppression function exists, mathematically it should be possible to extract it based on the first half of Eq. 2.29, given K and #. DFT calculations on a material such as Si can produce 0, and /- can be obtained through experimental measurements over a range of thermal length scales. The prospect of extracting S from experimental measurements is inviting, due to the inherent difficulty in calculating S by solving the BTE. For this experimental extraction scheme to work, S must be reformulated as a monotonically increasing function, and as an array rather than a matrix. A change of integration variable from A to 77, where now q is defined as = j(d) S(n)#(A) (-) = d/A leads to dq. (2.31) Here d is the thermal length scale in the measurement, A is the phonon mean free path, /-, is the normalized effective thermal conductivity, and ductivity per mean free path. Defining 7 as # is the thermal con- d/A makes S an increasing function of n. This reformulation makes it possible to treat S as the unknown array in the convex optimization, which uses the constraints that S increases smoothly and monotonically from 0 to 1. The reconstruction of S is complicated by the sharp features in O(A), which require a fine spacing in A for proper integration. O(A) for Si at 300 K is plotted in Fig. 2-20(a) [46]. The thermal conductivity accumulation function <b(A), shown in Fig. 2-20(b), is related to O(A) through <b(A) = f #(A')dA'. To test the viability of reconstructing S based on experimental / and known #, we consider the measurement on Si membranes discussed in Section 2.3.3. Si membranes form a nice test case for several reasons: DFT calculations for # Fig. 2-20(a)); we have a wide length scale range of measured for Si are available (see K data (see Fig. 2-17); and the correct suppression function is known to be given by the Fuchs-Sondheimer relationship (see Eq. (2.30)), so the quality of our S reconstruction can be evaluated. 58 x 7 -6- 10, 3 20.8 6- U S0.6 4-6 0.40 0 0 2- . 0.2 E L_(D C 10~9 10 10 10 A [M] 10 5 10~- 10 10 10 10~- 10-5 104 A [M] (b) (a) Figure 2-20: (a) Thermal conductivity per MFP and (b) thermal conductivity accumulation function for Si at 300 K from DFT calculations [46]. We proceed to reconstruct S for thin membranes with diffusely scattering boundaries, as shown in Fig. 2-21. To properly integrate 0, we used 1600 elements in the array of S, and we verified that this discretization could reproduce <b. The large number of elements in S required a higher smoothing factor to produce a smooth curve. We used a smoothing factor of 100. The resulting reconstructed suppression function shown in Fig. 2-21 differs somewhat from the exact Fuchs-Sondheimer suppression function, but can still effectively reconstruct the Si thermal conductivity accumulation function from measured Si membrane data. Determining suppression functions in this manner, while attractive due to the difficultly of solving the BTE, requires further study to verify viability. For example, the choice of smoothing factor influences the result due to the large length of the S array required to integrate 2.4 # reliably. Summary and future directions We have presented gray-medium and spectral solutions to the one-dimensional phonon BTE corresponding to the spatially-sinusoidal temperature profile in a TTG experiment. Our gray-medium analysis yielded an analytical solution that approached 59 1 1 [***~.~** CD) - 0.8 - S0.6 C - 0.4 ~0. 0.2 C/) --- reconst. 1 ~.-SFS C-4 -3 -2 -1 0 1 2 .3 10 10 10 10 10 10 10 10 ,q = d/A 4 10 Figure 2-21: Reconstructed heat flux suppression function for thin, diffusely scattering membranes (dashed line). The exact solution given by the Fuchs-Sondheimer relationship (see Eq. (2.30)) is shown for comparison (solid line). the diffusive limit for grating periods that were large compared to the gray-medium phonon MFP, and approached the ballistic limit for small grating periods. Spec- tral BTE solutions were found for Si and PbSe at 300 K using phonon dispersions and lifetimes for all six phonon branches from DFT calculations. We compared the spectral BTE decays to several approximate models: a single-MFP BTE solution, a frequency-integrated gray-medium BTE model, and a two-fluid model from Ref. [8] that combines the BTE with the diffusion equation. We found that the spectral BTE results for Si were well reproduced by the two-fluid model from Ref. [8], and that PbSe was reasonably modeled using our proposed frequency-integrated gray-medium BTE approach. We also showed that the contribution of ballistic phonons is small even for large reductions in keff compared to kbu1k. We went on to consider the inverse problem of reconstructing thermal conductivity accumulation functions from measured effective thermal conductivities and modeled suppression functions. While the suppression function from Ref. [8] produced better results for Si, the suppression function from our frequency-integrated gray-medium BTE approach produced reasonable results, and, in fact, worked better for PbSe. We anticipate that the latter approach, applied to different experimental geometries, 60 may offer reasonable estimations for modeling non-diffusive thermal transport, and extracting phonon spectral information from experimental measurements. Additionally, we demonstrated that the Fuchs-Sondheimer relationship could be used as a suppression function for reconstructing the thermal conductivity accumulation function for bulk Si from TTG measurements on Si membranes. Finally we explored the viability of deriving heat flux suppression functions experimentally, using the Si membrane data as an ideal test case, and found that such an approach is complicated by the sharp features in the thermal conductivity per MFP function. Our theoretical work demonstrated that TTG can be a useful tool for studying non-diffusive thermal transport. However, to study low thermal conductivity materials with short phonon MFPs, it is necessary to generate thermal length scales smaller than the diffraction limit of optical light. Membranes are one approach for generating small thermal length scales, but while Si membranes are commonly produced, thin membranes of other more exotic materials are challenging to fabricate. These factors motivated us to consider alternate methods for experimentally generating and measuring heat transport over small thermal length scales. 61 62 Chapter 3 Investigation of non-diffusive conduction with microfabricated wire grid polarizers Prior works have made progress in developing methods to observe non-diffusive heat conduction, but have been restricted in thermal length scale by the optical diffraction limit [4, 19], or have been limited to transparent substrates [3, 57, 58]. Diffraction limited methods have used laser diameter [4] or transient thermal grating period [19] as variable heating scales. Microfabticated metal heaters patterned on optically transparent substrates have been used to go beyond the diffraction limit [3, 57, 58]. In addition, thin Si membranes have been used [59, 54] as discussed in Section 2.3.3, but fabricating thin membranes of other materials is challenging. To study non-diffusive transport in generic opaque materials with short phonon MFPs, we require a method that can achieve small thermal length scales and detect the resulting heat transport. In this chapter, we explore one such method that uses a microfabricated wire grid linear polarizer on the surface of a generic sample of interest in conjunction with TDTR measurements. optical excitation of the underlying sample. 63 The polarizer is designed to minimize 3.1 Concept and design criteria for wire grid polarizer To achieve length scales below the diffraction limit of visible light and to minimize direct optical excitation of the sample under study, we explored using a wire grid linear polarizer (LP) fabricated on the surface of the sample, as illustrated in Fig. 3-1. Timedomain thermoreflectance, which was described in Section 1.3.2, is then performed with linearly polarized pump and probe beams. When the electric field directions of the laser beams are perpendicular to the transmission axis of the linear polarizer, a minimal amount of optical intensity will be transmitted into the underlying sample. The LP transmission axis is perpendicular to the metal wires, as indicated in Fig. 3-1. When the electric field direction is aligned parallel to the metal wires, free electrons in the metal act to absorb the light, but when the electric field is perpendicular to the lines where electrons are not as free to move, a maximum amount of light is transmitted through the grating. The metal wires act both as localized heaters and as transducers for detecting the change in thermoreflectance resulting from transient temperature relaxation as the delay time between the pump and probe beams is varied. The thermal length scale in the measurement will depend on the grating line width and period. We use COMSOL electromagnetic wave simulation software to design the geometry of the ID metal grating that acts to minimize light transmission into the underlying substrate. We have several design criteria. The grating must minimize transmission of the 800 nm probe and 400 nm pump beams into the underlying substrate. We also desire a structure that has a high likelihood of observing non-diffusive behavior, so we prefer small thermal length scales. Since the thermal length depends both on the grating line width and period, we desire small line widths with large periods to minimize the spreading of heat between neighboring lines. The grating should also be reasonable to fabricate. A schematic of the two-dimensional (2D) finite element simulation space is shown in Fig. 3-2, which consists of a single grating period L. Periodic boundary conditions are used to simulate an infinite 1D grating. 64 The top view: generic sample Al side view: transmission axis of LP E, 400 nm pump E, 800 nm probe Figure 3-1: Conceptual illustration of a ID metal grating acting as a linear polarizer (LP) to block pump and probe light from directly exciting a generic substrate. For minimum transmission through the grating polarizer, the pump and probe are linearly polarized with their electric fields aligned perpendicular to the polarizer's transmission axis. electric field of the light is aligned along the metal wires, perpendicular to the grating transmission axis, and the light propagates from the air towards the substrate. Figure 3-3 shows the transmittance and reflectance results for wavelengths A of 400 nm and 800 nm for various grating geometries, and provides some intuition about the geometries that are more effective at light blocking. For these calculations, the substrate material is A12 0 3 and the metal is Al. Shorter period (smaller L), higher aspect ratio (larger h) gratings have lower transmittance, and longer wavelengths are more easily blocked than shorter wavelengths. The range of grating line widths d with low transmittance is the largest for Fig. 3-3d, which combines small periods with high aspect ratios. A large range of d with low transmittance can be achieved by keeping the gap between the metal lines, L - d, constant and smaller than the wavelength, as shown in Fig. 3-4, which uses a 100 nm gap width. 65 Even though this constant small L air 3L Figure 3-2: Schematic illustrating the domain used for COMSOL calculations. gap approach achieves small optical transmission, it may be difficult to observe nondiffusive transport at larger values of d. The thermal length scale depends both on the period and the line width. For larger filling fractions, which correspond to larger ratios of d/L, non-diffusive effects are less apparent [60]. For interpreting measured non-diffusive effects, it is intuitive to keep the filling fraction constant [60]. Figure 3-5 shows transmittance and reflectance results where the filling fraction is held constant at L = xd, where x is an integer. Only calculations for A = 400 nm, which will have higher transmission than A = 800 nm, are shown. Larger values of x will result in less thermal information sharing between neighboring metal wires making the observation of non-diffusive transport more apparent, but smaller values of x result in improved light blocking. Figure 3-5(a) shows that a transmittance of less than 10% is possible for d as large as 100 nm for x = 2, and Fig. 3-5(b) shows that higher aspect ratio lines (larger h) further reduce the transmittance. Thus far, we have determined that L = 2d with h = 100 nm is a feasible design for minimizing transmission into the sample, and that h = 200 nm would be preferable if fabrication constrains allow. For h > 200 nm, in addition to fabrication challenges, there could be issues with TDTR signal-to-noise, whereby the measurement is not 66 L = 200 nm, h =100 nm (a) CD 8 (b) 1- U) CO C: E 0.6 W 0.4 0 0.6 - + C) L = 100 nm, h= 100 nm (D orA I W 0.4 C 04 0.2 as0. 2 4-1 0 -0 10 -9 -8 10 10 d (m) -7 10 -10 10 -9 0-8 d (m) U) 10 "" 1-7 1) (c) L = 200 nm, h = 200 nm (D (d) 1 1 C L = 100 nm, h = 200 nm E 0.8 - E 0.8.0.6 0 CU 0.4- M (D - 0.4 I-,Y 0 - c 0.2- 10 - 0 10 9OeE 0.6 10 10 d (m) " ~" S 10-7 0 -10 10 10 10 d (m) 107 Figure 3-3: COMSOL calculations for transmittance (solid lines) and reflectance (dashed lines) for 800 nm (square markers) and 400 nm (circle markers) light for various grating geometries. For these calculations, the electric fields are aligned in the direction of the metal wires, the substrate material is A12 0 3 and the metal is Al. 67 L-d= 100 nm, h =100 nm 8 100 CO E 10 C O 10 Ca () -2 0 10 Cz U 10- -0 10 d (m) Figure 3-4: COMSOL calculations for transmittance (solid lines) and reflectance (dashed lines) for 800 urn (square markers) and 400 nm (circle markers) light with a constraint on the relationship between L and d such that the gap between grating lines is held constant at L - d = 100 nm. For these calculations, the electric fields are aligned in the direction of the metal wires, the substrate material is A12 0 3 and the metal is Al. very sensitive to the thermal conductivity of the underlying substrate. The preceding calculations used Al as the metal, which is preferred due to its high thermorefeletance response at our 800 nm probe wavelength [61]. Figure 3-6 shows some transmittance results for different metals. Since Al has a lower or comparably low transmittance to other metals, in conjunction with its favorable thermoreflectance response, we choose Al as the metal for our grating design. We also verify our transmittance calculations using Bi 2 Te3 as the substrate material, as shown in Fig. 3-7. 3.2 Microfabrication In Section 3.1 we determined that a ID Al grating with a thickness of 100-200 nm, a filling fraction of 50%, and a line width of 100 nm or less, would have lower than 10% transmittance. To fabricate these structures, we use a combination of interference lithography and reactive ion etching. Interference lithography is a convenient method for rapidly patterning large areas, and reactive ion etching can produce high aspect ratio structures with vertical sidewalls. 68 L = xd, h = 100 nm .. w. 10 10 x =4 10 x=3 10x =2 10 --8 10 d (m) (a) 10 L = xd, h = 200 nm -- ------ 10 10 -- -- -- +- 10 10-1 10 10 10 x=2 1 ) --8 d (i) 10 (b) Figure 3-5: COMSOL calculations for transmittance (solid lines) and reflectance (dashed lines) for 400 nm light with a constraint on the relationship between L and d such that the filling fraction is kept constant with L = xd, where x is an integer. Grating line heights of (a) 100 nm and (b) 200 nm are considered. For these calculations, the electric fields are aligned in the direction of the metal wires, the substrate material is A1 2 0 3 and the metal is Al. Figure 3-8 illustrates the basic process flow, and Fig. 3-9 shows some scanning electron microscope (SEM) images of various steps along the process flow. Fabrication was carried out in the NanoStructures Laboratory (NSL) at MIT, with extensive process development consultation from NSL personnel. The starting wafer substrate is coated in Al followed by SiO 2 using electron beam evaporation. The SiO 2 will serve 69 L = 2d, h = 100 nm Ag . 10C -3 10 E Cr 10 10 Al 108 d (m) 10- . Figure 3-6: COMSOL transmittance calculations for 400 nm light, varying the type of metal used for the one-dimensional grating. For these calculations, the electric fields are aligned in the direction of the metal wires and the substrate material is A12 03 = 2d, h =100 nm )L C 100................... C', - E 10 C (', 0 c CD 10 -1 -2 10 -4 10 10 d (m) Figure 3-7: COMSOL transmittance (solid line) and reflectance (dashed line) calculations for a Bi 2 Te 3 substrate patterned with a one-dimensional Al grating, using 400 nm light with an electric field in the direction of the Al wires. as a mask for etching the Al. An antireflection coating (ARC), which prevents back reflections during the lithography process, is spun on. A thin SiO 2 layer is evaporated on the ARC to serve as an etch mask. Since SiO 2 is a hard mask, a thicker layer of ARC may be etched than if photoresist alone served as the mask. Photoresist (PR) is spun on, and the wafer is baked prior to exposure. 70 The PR is exposed with a Lloyd's mirror interference lithography (IL) system [62] using a 325 nm laser source. The patterned period is set by the angle of the mirror, which divides the source into two overlapping coherent beams. The source is also expanded and made to have a uniform intensity over the exposure area, which can be as large as a 3 in wafer, using a lens and pinhole assembly. After exposure, the wafer is baked to harden the resist before development. The resulting developed resist is shown in Fig. 3-9(a). The developed resist is used as the first mask in a multi-step reactive ion etching (RIE) process. RIE is a dry etch process that directs ions in an energetic plasma towards the wafer plate using a large voltage difference. The etch process occurs through a combination of ion bombardment, which mechanically removes material, and chemical reactions, which speed material removal. The SiO 2 layers are etched with a CF 4 plasma, and the ARC is etched with an 02 plasma. The resulting SiO 2 to be used for Al RIE is shown in Fig.3-9(b). RIE of Al is a challenging process, and is discussed further in Section 3.2.1. After Al RIE, the remaining SiO 2 mask shown in Fig. 3-9(c) needs to be removed. This is done by spinning a thick layer of ARC that covers the grating structures, as shown in Fig.3-9(d). Another RIE process is used, etching the ARC, SiO 2 and ARC, to produce the finished Al grating shown in Fig.3-9(e). ARC IL RIE spin RIE Figure 3-8: One-dimensional grating fabrication process flow. 3.2.1 Dry etching Al RIE of Al is a challenging process that had not previously been developed at MIT's NSL. Through systematic iteration, we developed a reliable process for etching Al in 71 (a) PR Si02 ARC Si02 Al Si (b) ~I (c) (d) (e) Figure 3-9: SEM images of the 1D grating fabrication process: (a) post IL exposure and PR development, (b) the SiO 2 mask before Al RIE and (c) post Al RIE, (d) spun ARC for SiO 2 mask removal, and (e) the completed grating structures. 72 the NSL's inductively coupled plasma (ICP) RIE. Our recipe uses equal amounts of C1 2 and N 2 gases at a low pressure (1.6 mT) with a high RF bias (300 W) and little or no ICP coil power. C12 reacts chemically with Al to form AlCl 3 , even in the absence of a plasma. Low pressure helps promote the removal of AlCl 3 products and improves the anisotropy of the etch. The C12/N 2 plasma is sparked at a high pressure (15 mT) for 5 seconds, and subsequently lowered to 1.6 mT, where the remaining etch time for a 100 nm thick Al layer is -2 min. The electrostatic chuck holding the carrier wafer during the etch is maintained at 35 C. After removing etched Al from the vacuum chamber, the Al undergoes a post-etch corrosion process upon exposure to the atmosphere since adhered molecules of AC1 3 undergo hydrolysis, producing HCl. Some examples of post-etch corrosion are shown in Fig. 3-10. To prevent this corrosion process, we developed a two step post-etch treatment. After Al etching, but before venting the vacuum chamber, a high density, low RF bias N 2 plasma is sparked and allowed to bombard the sample for 5 mins. This helps to remove adhered molecules of Al etch products. Immediately after venting the vacuum chamber, the sample is immersed in deionized (DI) water, and then rinsed in DI water for ~2 mins. This post-etch treatment prevented any observable corrosion over the course of weeks following Al etching. 3.2.2 Polishing polycrystalline Bi 2 Te 3 samples Bi 2Te 3 is a promising thermoelectric material due to favorable electrical properties and low thermal conductivity. The thermal conductivity can be reduced by nanostructuring [63]. We wanted to investigate whether we could observe any size-dependent thermal conductivity using our wire grid polarizer approach. To do so, we had to fabricate ID Al grating structures on the surface of polycrystalline Bi 2Te 3 samples. Polycrystalline Bi 2 Te 3 samples were provided to us by Professor Zhifeng Ren from the University of Houston. These were fabricated from powdered forms of Bi 2Te 3 that were hot pressed into 2 cm diameter cylinders with grain sizes of tens of nanometers. The cylinders were sliced into 1-2 mm thick pieces with a diamond saw. These small wafers had to be polished before we could microfabricate our grating structures. 73 (a) (b) Figure 3-10: Examples of post-etch corrosion (a) where corrosion products form clusters and (b) where corrosion results in removed sections of Al. A polishing process for polycrystalline Bi 2Te 3 had been developed previously in our group [64], but required multiple complex steps and proprietary chemicals. We developed a simpler process that reliably achieves 2 nm RMS surface roughnesses with commercially available polishing products. We use a commercial benchtop motorized polishing machine (South Bay Technology model 920 lapping and polishing machine) fitted with a central rotary plate ("wheel") and a yoke ("arm") for additionally rotating the sample holder. Starting with a rough, as-diced polycrystalline Bi 2Te 3 wafer, we use a 6 Mim diamond or A1 2 0 3 suspension on a hard polishing cloth (Buehler Trident or TexMet C) to planarize the surface, followed by a 1 pim A12 0 74 3 suspension to reduce the surface roughness. Polishing suspensions are introduced at a rate of -1-2 drops/min, and DI water at -1-2 drops/sec is used for lubrication. For this initial planarizing and smoothing, the polishing wheel and arm speeds are fast and contra rotating. The planarizing step takes -2-5 mins and the rough smoothing step takes -5-10 mins. Between each polishing step, the sample is throughly rinsed with DI water. Fine polishing is achieved in two subsequent steps. diluting commercially available Buehler 50 nm A1 2 0 the solution has a pH of -7.5. 3 A suspension is made by solution with DI water until The diluted suspension is introduced at a rate of -1-2 drops/sec. A hard polishing cloth (Buehler Trident) is used, with fast wheel and arm speeds that are counter rotating. After -10-15 mins, the sample has a mirror finish with some visible uniform shallow scratches. The final polishing step uses the diluted 50 nm A12 0 3 suspension introduced at a rate of -2 drops/s, and a soft polishing cloth (Buehler Microcloth) with fast wheel and arm speeds that are counter rotating. A 250 gram weight is added to provide more downward force on the sample. The total time for this step is critical, and -30 s. Polishing too long will result in pits from corrosion or an orange-peel texture on the surface, and polishing for too short a time will not remove all the scratches. After -30 s, a copious amount of DI water is introduced for -15 s while spinning polishing wheel and sample are still in contact, which helps to remove polishing slurry from the sample surface. After this step, the sample is cleaned in warm 45'C DI water in an ultrasonic cleaner to remove any remaining polishing debris. The finished sample has a mirror finish, and atomic force microscope (AFM) measurements over multiple 100 pm 2 measurement areas indicate an RMS surface roughness of -2 nm, as shown in Fig. 3-11. 3.2.3 Resulting one-dimensional grating structures Examples of final fabricated gratings are shown in Fig. 3-12. Figure 3-12(a) shows a thinner line width grating with a higher aspect ratio than Fig. 3-12(b). These grating - structures were fabricated on wafers of Si, fused silica, and polycrystalline Bi2 Te3 Subsequent sections describe our TDTR measurement results on these structures. 75 10 nm 5 nm __ 0 nm 10 pm Figure 3-11: Representative AFM measurement of a polished polycrystalline Bi 2 Te 3 sample. The RMS surface roughness over the imaged 100 pm 2 area is 1.6 nm. Measurements over several locations on the sample showed comparable roughness. AFM scan courtesy of Lingping Zeng. 3.3 Optical transmission results To evaluate the polarizing efficiency of our fabricated gratings, we measure the extinction ratios of the gratings on fused silica substrates, which are transparent to the wavelengths in our TDTR system. A schematic for the transmission measurement setup is shown in Fig. 3-13. The linearly polarized pump and probe beams are circularly polarized by a quarter-wave plate (A/4) at 450 , and then linearly polarized by a linear polarizer (LP) in a rotational stage. The pump and probe beams are linearly polarized in the same direction, and focused onto the fused silica sample with the patterned Al grating structure. The focusing angle for the beams is < 4'. The transmitted light intensity is collected by a detector. By rotating the LP, the angle between the electric fields of the pump and probe beams and the grating transmission axis, 6, can be varied. 6 transmission Tmin, and 0 = = 900 gives the minimum 00 gives the maximum transmission Tmax. Figure 3-14 76 (a) (b) Figure 3-12: Examples of fabricated 1D Al gratings. (a) Grating with a 200 nm pitch and an 85 nm line width, that is 160 nm thick. (b) Grating with a 200 nm pitch and a 105 nm line width, that is 85 nm thick. plots the ratio of the transmitted intensity I to the incident intensity 1, as a function of 0, assuming an ideal linearly polarizing grating. Measured extinction ratios, Tmin/Tma, for the 400 nm pump beam and 800 nm probe beam for various grating geometries are shown in Table 3.1. The extinction ratio for the 85 nm line width, 85 nm thick grating was > 10% for 400 nm light, and increasing the aspect ratio by increasing the thickness to 160 nm improved the extinction ratio substantially. Low extinction ratios indicate that our fabricated gratings are indeed acting as reasonable polarizers. For comparison, COMSOL calculated 77 pump probe 1 .j PIN detector lens sample lens LP. in M/4 cold rot. stage mirror B.S Figure 3-13: Transmission measurement setup where either the 400 nm pump or the 800 nm probe is used. 1 1 D grating - 1Dg raing linear polanizer 0.8 0.6 0 0 o 0.4 Qo. aarSM xIs- 1/0 0.2 0 10 20 30 40 50 0 (deg.) 60 70 80 90 Figure 3-14: Idealized linear polarizer transmitted intensity as a function of angle. extinction ratios are shown in Table 3.2 for the same grating geometries as those in Table 3.1. The fabricated gratings (see Table 3.1) have extinction ratios 1.3 to 4 times greater than the idealized geometries assumed in our COMSOL calculations (see Table 3.2). 3.4 TDTR results For reflection measurements, we use the setup illustrated in Fig. 3-15, which is similar to our transmission setup illustrated in Fig. 3-13, with the exception of the detector 78 Table 3.1: Measured extinction ratio, Tmin/Tma, for various fabricated gratings for the 400 nm pump and 800 nm probe beams. All gratings have a period of 200 nm and were fabricated on substrates of fused silica. grating geometry 800 nm 400 nm line width 105 nm, thickness 85 nm line width 85 nm, thickness 85 nm line width 85 nm, thickness 160 nm 1% 1.7% 0.2% 4.8% 16.6% 3.6% Table 3.2: Extinction ratios, Tmin/Tmax, from COMSOL calculations corresponding to the fabricated grating geometries in Table 3.1. grating geometry 800 nm 400 nm line width 105 nm, thickness 85 nm line width 85 nm, thickness 85 urn line width 85 mu, thickness 160 nm 0.5% 1.3% 0.06% 3.6% 9.1% 0.9% position and an added color filter for preventing pump light from entering the detector. This is the same as a standard TDTR setup, like that discussed in Section 1.3.2, with the addition of a linear polarizer in a rotation stage. The pump beam is modulated, and the reflected probe beam is collected. The signal from the detector is mixed with the reference modulation profile in a lock-in amplifier, and the resulting in-phase and out-of-phase amplitudes for a given delay time are evaluated with a thermal model to extract model parameters of interest. In our experiment, the unknown parameter of interest is the substrate thermal conductivity. pump sample lens pro be L.P. in k/4 cold rot. stage mirror B .S. lens blue PIN filter detector Figure 3-15: TDTR grating measurement setup. 79 3.4.1 Grating heat transfer model We introduced the heat transfer model for TDTR in Section 1.3.3. Modeling the case where the transducer is a 1D grating requires a slight modification. We follow the approach of Minnich [57], who developed a model for a transducer formed from a 2D grating. We assume that the metal layer cannot conduct heat in the in-plane xy-direction, and can only conduct in the cross-plane z-direction. This approach also assumes infinite pump and probe beam diameters, and neglects any Gaussian intensity variation. The heat diffusion equation solution in Section 1.3.3 assumes cylindrical symmetry, but for the ID grating geometry, it is more convenient to solve the heat diffusion equation in Cartesian coordinates and to use a spatial Fourier transform instead of a Hankel transform. The solution takes the form of Eq. (1.5) with IM 2 = k.,Y2 + iWC . (3.1) Here kxy is the in-plane thermal conductivity which is considered isotropic, k, is the cross-plane thermal conductivity, sx is the x-direction spatial Fourier transform variable, and w is the radial frequency temporal Fourier transform variable. The time harmonic heating at the top surface qtop has a spatial profile of a ID square wave, assuming that only the grating wires are heated. The spatial square wave is represented by a Fourier series. In the spatial and temporal transform domains the top surface heat flux qtop becomes qtOP(Q) = q,6(w - w,) E an6(Q - n ), (3.2) n=-oo with Fourier series coefficients an of an if n 0, sin(nrQd/2)/(n7r) if n 7 0, d/L (3.3) where the spatial frequency is Q, = 27r/L. Here L is the 1D grating period, d is the 80 grating line width, w, is the time-harmonic excitation frequency, and q is the pump intensity. The top surface temperature (see Eq. (1.9)) is weighted by the same spatial square wave probe profile, and the resulting frequency response h(w) is given by [57] h(W) cX E |X| 3.4.2 . C-D (3.4) n n Measurements varying pump laser diameter Our thermal model assumes an infinite laser heating size, and as a result, a large enough pump size is necessary to achieve reasonably accurate results. Figure 3-16 shows fitted measurement results for a 105 nm line width, 200 nm period gratings on fused silica and Si substrates. For the fused silica substrate, which has a low thermal conductivity, we observe a dependence on the pump laser spot size, which we attribute to the large heating diameter assumption inherent in our thermal model. At small pump sizes, heat spreading along the metal wires, which is not captured in our thermal model, becomes appreciable. At sufficiently large pump diameters, the measured substrate thermal conductivity asymptotes to the bulk value of the thermal conductivity of fused silica. Measurements on a Si substrate are not sensitive to pump spot size, because Si has a higher substrate thermal conductivity. The measured Si thermal conductivity, however, is much lower than the bulk value, indicating non-diffusive transport. At a thermal length scale of 105 nm, we expect to see size effects in Si based on the distribution of heat-carrying phonon MFPs [48, 46]. 3.4.3 Measurements varying angle between laser polarization and grating transmission axis Varying the angle between the laser polarization and the grating transmission axis influences the transmitted intensity, as shown in Fig. 3-14. The minimum transmission occurs at 0 = 900 and the maximum transmission occurs at 0 = 00. Our reflection measurement setup (see Fig. 3-15) allows us to easily vary 0 by rotating the linear 81 50 ,40I - 1.5 - 30 1-3 _020- _0 C C -Fz0.5 -C E E 10- a) a) 80 100 120 140 pump diameter ([tm) 90 160 (a) 90 100 110 120 pump diameter ([rm) 130 (b) Figure 3-16: Measured substrate thermal conductivity as a function of pump laser diameter for (a) fused silica and (b) Si. ID Al gratings with 200 nm periods, 105 nm line widths, and 85 nm thicknesses were fabricated on each substrate, and TDTR measurements were performed with linearly polarized pump and probe beams aligned perpendicular to the grating transmission axes (see Fig. 3-15). polarizer. The pump and probe beams are polarized in the same direction. The measured reflectance signal from the lock-in amplifier is shown in Fig. 3-17 for 0 = 00, 450 and 90'. The in-phase x, and out-of-phase y, lock-in outputs are combined to form amplitude R, and phase #, data as a function of delay time T between the # pump and probe pulses. R has relative units, and can be scaled arbitrarily, while has absolute units of angle. For the fused silica substrate, the phase signal is constant regardless of 0, and the amplitude signal is constant except for an offset, as shown in Fig. 3-17(b). The oscillations in the 0 = 0' and 45' curves result from surface acoustic waves. Some probe light is reflected off the surface of the fused silica substrate and interferes with probe probe light reflected off the surface of the metal grating. As the substrate and grating thermally expand, the interference can be constructive or destructive, resulting in oscillations in the measured signal on top of the thermal decay profile. Since fused silica is transparent to the pump and probe wavelengths, the measured thermal response should be the same regardless of 0. For Si, pump absorption in the substrate will influence the heating profile, and the 82 -00 -- ..... 90* -0 .0 -- 45* 510 -9Q* -- - 100 450 - r --------- ----- - -1.. . . . . 10 -20 100 50 -40 --- 60 - -50 -100 0 1 2 3 -(ns) 4 5 2 1 0 6 3 t (ns) 4 5 6 (b) (a) Figure 3-17: Reflection results for various angles 0 on (a) Si and (b) fused silica showing both amplitude R and phase q TDTR data as a function of delay time T. probe reflection from the substrate will influence the reflectance signal. As 0 is varied, we observe dramatically different R and # signals from the Si substrate sample, as shown in Fig. 3-17(a). The minimum transmission angle (0 = 90') produces a familiar thermal decay profile, while the maximum transmission angle (9 = 0) produces a much faster R decay, more indicative of bare Si. For reference, the TDTR response from bare Si with no transducer layer is shown in Fig. 3-18. In Fig. 3-19, we examine the TDTR signal from a range of angles on a Si substrate sample. Figure 3-19(a) shows angles near the maximum transmission angle, and Fig. 3-19(b) shows angles near the minimum transmission angle. Figure 3-20 more closely examines angles near 9 = 90', looking at variations of 5' and 15' on either side. We observe good agreement between 0 = 850 and 95' and between 9 = 750 and 105', providing confidence in our determination of 9 = 900. clear difference that 50 makes in the shape of R and 83 #. Furthermore, we note the Figure 3-14 shows that going 100 E- 10 0 -10 -* -15.-20-25 0 1 2 3 x (ns) 4 5 6 Figure 3-18: Thermoreflectance signal from bare Si. from 0 90' to 6 = 85' corresponds to less than 1% in transmitted intensity, which suggests that even a small amount of transmitted light influences the TDTR signal in an appreciable way. 3.4.4 Result summary The substrate thermal conductivity results obtained from fitting the thermoreflectance response measured at 0 = 900 for various samples are shown in Table 3.3. For fused silica and polycrystalline Bi 2 Te 3 substrates, no measurable deviation from bulk values 84 100 10 100 0 100. -20 50 50 a0 0 0 -50- -40- *15* -25 . -60- -- 35* -~-5* r --5 -80 .. 75* -Mo 201 -1001 0 - -100 1 2 3 -r (ns) 4 5 -- 0 6 1 2 3 -r (ns) 4 5 6 (b) (a) Figure 3-19: Reflection results for various angles on Si at (a) high transmission angles and (b) low transmission angles. is observed, suggesting diffusive transport behavior. From the Si substrate, we observe a clear deviation from the bulk thermal conductivity, but note that our results are likely influenced by some signal contribution from the Si substrate, so while a nondiffusive effect is likely, the precise fitted thermal conductivity values are unreliable. A lack of observed size effect in polycrystalline Bi 2 Te 3 suggests that the dominant phonon mean free paths (MFPs) are smaller than 100 nm in that material, which has been confirmed by recent density functional theory calculations that suggest that the heat-carrying phonon MFPs in crystalline Bi 2 Te 3 are < 10 nm [65]. 3.5 Future directions Since even a small amount of optical transmission into the underlying substrate con- tributes to experimental uncertainty, it may be advantageous to pursue an approach 85 -750 -- 85* -- 90" - -950 100 - -105* 10 - - - 0 -10-20C) 0, V -30-40-50 I -60 0 2 1 3 4 5 6 -c (ns) Figure 3-20: Reflection results for angles around the low transmission angle for a one-dimensional Al grating on Si. Table 3.3: Measured effective thermal conductivity in units of W/mK for substrates of fused silica, Si and polycrystalline Bi 2 Te 3 with two different Al grating geometries compared to the bulk substrate thermal conductivity. The 105 nm grating line width sample had a thickness of 85 nm, and the 85 nm line width sample had a thickness of 160 nm. line width 105 nm line width 85 nm kbu1k substrate fused silica Si 1.4 142 1.36 45 polycrystalline Bi 2 Te3 1.3 1.4 1.4 33 that guarantees no direct optical excitation of the substrate. One approach could be to fabricate metal gratings with a capping metal layer, as shown in Fig. 3-21. The 86 gratings are fabricated as before, and the gaps are filled with a low thermal conductivity polymer, like the antireflection coating material used for mask removal in Section 3.2. The structure is then coated in an optically thick layer of Al, which acts as a transducer for the TDTR measurement. Most of the heat should conduct through the Al capping layer to the Al wires, circumventing the low conductivity polymer areas. Properly interpreting the TDTR data would require a refined thermal model, but preliminary testing suggests that non-diffusive transport can be observed for Si substrates, and that diffusive transport is observed in fused silica substrates, as expected. Figure 3-22(a) shows fitted substrate thermal conductivities for structures like that shown in Fig. 3-21, where the pitch of the grating structures was 180 nm and the line widths were smaller than 80 nm. The model used for fitting was rather crude, treating the structure with three layers: an Al top layer, an interface layer consisting of the grating, polymer and bounding interfaces, and a substrate layer. The unknown fitting parameters were the interface layer resistance and the substrate thermal conductivity. In spite of crude modeling, we observe apparent non-diffusive transport in Si and diffusive transport in fused silica substrates. Figure 3-21: Concept for structures that prevent optical transmission into the substrate. The gaps between the Al grating lines are filled with polymer (ARC) as shown on the left, and then an Al capping layer is added as shown on the right. , The primary findings of this work are twofold. First, for polycrystalline Bi 2Te 3 length scales smaller than 100 nm are required to possibly observe non-diffusive ef- 87 102 E -------------------------10 . .......... .........74..... 72..... 74 72 76 78 . 0 .1007 0 linewidth (um) (a) -20 -40 (D - - -60 -Si -80[ 0 1 2 3 -r (ns) 4 5 ' 6 -100 S 0 71 nm' Si, 76. nm - fsed s76.5 a 71n -... fused silica, 71 nm -ue 1 -iia -6 2 4 3 -c (ns) - .0 10 5 6 (b) Figure 3-22: (a) Using a crude model, where the grating-polymer layer and bounding interfaces are treated as having some unknown lumped resistance value, we extract substrate thermal conductivity values for substrates of Si (open circles) and fused silica (open squares) patterned with grating structures like that pictured in Fig. 321. The dashed line shows the bulk thermal conductivity of Si and the dotted line shows the bulk thermal conductivity of fused silica. (b) For reference, raw TDTR phase data traces, 0(r), collected with a pump modulation frequency of 9 MHz, are shown for samples corresponding to the fits in (a). fects. Length scales as small as ~20 nm could be achieved with ebeam lithography methods. Another approach might be fabricating thin membranes of Bi 2Te 3 , where the membrane thickness would serve as the thermal length scale. The second finding of this study was that even small amounts of optical transmission (increases of 1%) into the substrate influence the TDTR signal. Optical transmission could be further reduced by increasing the wavelengths used in the measurement, for example, by implementing a two-tint TDTR approach that uses near infrared light for both 88 the pump and the probe [66]. To fully eliminate direct excitation uncertainty, the substrate would either have to be transparent to the wavelengths used, or some light blocking structure like that of the cap structure described above would be needed. 89 90 Chapter 4 Frequency-domain representation of TDTR data, and applications for studying non-diffusive conduction Since the study of phonons in low thermal conductivity materials such as thermoelectrics likely requires thermal length scales on the order of 10 nm or smaller, we are motivated to develop an alternate experimental approach. The challenges and cost associated with microfabrication at 10 nm length scales, which is the current state of the art, are prohibitive. Recent work [18] suggested that thermal penetration depth could serve as an effective means of probing small thermal length scales without the need for microfabrication. Periodic heating on a semi-infinite substrate results in a periodic temperature response that decays with depth. Solving the heat diffusion equation in one dimension, X, for a semi-infinite material subject to a heat flux input of q = q0 cxp(iwt) results in a transient temperature profile of T(x, t) = O k V/iw/a C"e--" eiw, which has a characteristic thermal penetration depth of 91 (4.1) k rCf dTPD - (4.2) Here, k is the substrate thermal conductivity, C is the volumetric heat capacity, a = k/C is the thermal diffusivity, q, is the magnitude and f is the frequency of the sinusoidal heat flux input, and w = 27rf. Figure 4-1 illustrates the transient temperature profile, along with the thermal penetration depth. At dTPD, the temperature envelope has decayed to ~37% of the surface value. At 2 dTPD, the temperature envelope will have decayed to -13.5% of the surface value. T(x,t)1/f qo To ILL t -AdD X Figure 4-1: Illustration of thermal profiles, T(x, t), in a semi-infinite solid that result from excitation by a sinusoidally periodic heat flux q(t). The dashed line shows the definition for the thermal penetration depth dTPD given in Eq. (4.2). The thermal penetration depth decreases with an increasing frequency of the periodic heat flux. As a first order estimate for the frequencies required to achieve 10 nm length scales in various materials, Fig. 4-2 shows dTPD(f) from Eq. (4.2) for three substrate materials which span a range of thermal diffusivities: Si, A12 0 3 , and fused silica. Figure 4-2 suggests that frequencies as high as 3 GHz are needed to achieve 10 nm thermal penetration depths in low thermal conductivity materials like fused silica. Experimentally generating and measuring such high frequency responses is challenging, but recent advances in optical-thermal measurement systems have made progress [18, 24]. Frequency-domain thermoreflectance (FDTR), which we introduced in Section 92 104 -Si - -A 2 O3 --. glass 103 E 10 2 10 100 102 101 3 10 4 f (MHz) Figure 4-2: Thermal penetration depth dTPD for semi-infinite solid slabs of Si, A12 0 3 and fused silica heated by a sinusoidal heat flux of frequency f (see Eq. (4.2)). 1.3.1, uses a sinusoidally modulated continuous wave (CW) pump beam to produce time-harmonic heating resulting in surface temperature oscillations at the modulation frequency, which are monitored by a CW probe beam. The modulation frequency in FDTR typically varies from kHz to -10 MHz [21, 22, 23]. Recently, an extension of the frequency range up to 200 MHz was reported, along with a reduction in the measured effective thermal conductivity of Si at high frequencies [18, 24]. Unlike FDTR, time-domain thermoreflectance (TDTR), which we introduced in Section 1.3.2, exhibits heating that is not time-harmonic but is comprised of many frequency components. The frequency content of TDTR measurements includes high frequency information, principally limited only by the laser pulse duration (typically ~200 ps). It would be advantageous if frequency components of the TDTR response could be separated and represented in a form similar to FDTR, i.e., in terms of the amplitude and phase of the surface temperature response to time-harmonic heating. We demonstrate that such a representation is indeed possible, and we extract frequency data of up to 1 THz. Our method not only allows a direct comparison of TDTR and FDTR data, but also enables measurements at high frequencies currently 93 not accessible to FDTR. 4.1 Theoretical foundation We briefly review the signal formation in FDTR and TDTR and we demonstrate that the frequency components in a TDTR signal can be separated and represented in a form identical to FDTR, i.e., in terms of the amplitude and phase of the surface temperature response to time-harmonic heating. 1/f0 1/10 1/f (a) -+ )+- (b) Figure 4-3: Typical laser heating profiles for (a) FDTR and (b) TDTR measurements. 4.1.1 FDTR In FDTR, the heat input supplied by the pump laser is a simple sinusoid with an angular frequency w, = 27rf0 , as illustrated in Fig. 4-3(a), which may be expressed as q(t) = qe"Aot, (4.3) where q, is the pump power modulation amplitude. This periodic heat input results in surface temperature oscillations with the same periodicity, 0(t) = qoe"wth(wo). (4.4) While the frequency is set by the pump, the magnitude and phase of the surface tem94 perature with respect to the pump are determined by sample properties. The complex function h(w) describes the frequency-domain response of the sample, and depends on parameters such as the substrate thermal conductivity and interface conductance. In practice, often only the phase of h is measured as w, is varied, because accurate measurements of the magnitude are more difficult [18]. h is typically modeled by solving the heat diffusion equation to find the sample's transient surface temperature response to a sinusoidal heat input [28, 29, 31], which is described Section 1.3.3. In FDTR, sweeping through all values of w, is necessary to fully determine h(w). 4.1.2 Single-shot TDTR If instead of time-harmonic excitation, the sample were heated by a short laser pulse approximating a delta function, the surface temperature would be described by the time-domain response, h(t). In the linear regime,' h is related to the frequency- domain response h by a Fourier transform, h(w) = f h(t) exp(-iwt)dt. By measuring h(t), one would fully determine h(w). To implement the single pulse excitation in an experiment, one would need to use a low laser pulse repetition rate to make sure that h decays to zero before the next laser pulse strikes the sample. In such an experiment, which could be called "single-shot TDTR," h, and consequently h, could be determined in a single measurement, provided that the signal-to-noise ratio was high enough. TDTR is typically performed at a high repetition rate (-80 MHz), which improves signal-to-noise, but results in accumulative effects of multiple pulses that complicate the data analysis. Practical difficulties would need to be overcome to implement a single-shot TDTR measurement. The laser repetition rate would need to be < 80 MHz (40 MHz would likely suffice), and the delay length required to capture the full transient temperature IThe linear response model normally used in the analysis of TDTR and FDTR measurements [28, 29, 31] is valid as long as temperature variations are small compared to the background temperature. In TDTR, this assumption may be inaccurate at early times (typically < 1 ps) when the non-equilibrium electron temperature rise in the metal film may be significant even for a moderate excitation fluence [67, 68]. We assume that the linear response model holds for slower dynamics determining frequency components of the thermoreflectance response below 1 GHz. However, establishing the domain of validity of the linear response model in TDTR requires further investigation. 95 response would be long (7.5 m in the case of a 40 MHz repetition rate). Aligning long movable delay lines is challenging, as it is necessary to prevent the beam from walking or changing diameter as the delay line is swept. Signal-to-noise would also be a challenge, requiring averaging at each delay position. TDTR is more easily implemented than single-shot TDTR, but the data analysis is more involved. 4.1.3 TDTR The analysis of TDTR data requires treating the accumulative effects of multiple pulse excitations. The heat input supplied by a TDTR pump beam can be approximated by a train of delta pulses modulated by a sinusoid as illustrated in Fig. 4-3(b), which can be expressed as q(t) = Qoe i~j~t 0 t -27k(45 --- oo , (4.5) k=-co where Q, is the absorbed pump pulse energy and the period between pulses is 2wr/ws. To convert this transient heat flux into a frequency-domain expression, we utilize the fact that q(t) is a product between an impulse train and a sinusoidal modulation function. 6( ) E 0 First, we consider the Fourier transform of the impulse train: 2 correin the 6time-domain S (w - kwo). A product 0_0 sponds to a convolution in the frequency-domain, and the special case of a product with a sinusoidal function in the time-domain leads to a shift in the frequency-domain: exp(iwet)x(t) -*(w - w,). Thus, in the frequency-domain, Eq. (4.5) becomes 00 4(w Q~~ E 6(w - kw, - w,,). (4.6) The surface temperature 6(w) relates to the heat input by 0(w) = h(w)d(w), hence O(w) = h(w)Qow, 6(w - kw - w,). (4.7) k=-oo 2 Note that the summation from k = -oo to oo is important, because the Fourier transform of an impulse train is derived using a Fourier series representation, which depends on having a periodic function. 96 In TDTR measurements, the probe is derived from the same laser as the pump, and can be treated as a train of delta pulses at the laser repetition frequency, delayed relative to the pump pulses by a delay time r. The incident probe power is given by qi t) =Q, >] M=00 DO t -27rm-T(48 6 (t -2-rr , (4.8) which in the frequency-domain becomes Vj(w) = Qiw (4.9) 6(w - mws)e-i""r, E m=-OC where Qj is the probe pulse energy. Equation (4.9) is derived using the theorem that a shift in the time-domain leads to a modulation in the frequency-domain: x(t - t,) <exp(-iwt0 )z(w). The reflected probe power, which will be collected by the detector, is given by the product of the sample's thermoreflectance response and the incident probe power, q,(t) = Cth(t)qi(t), where Cet is the thermoreflectance coefficient which relates the change in reflectance to the change in temperature. The power of the reflected probe beam is assumed to vary proportionally to the change in surface temperature, which is a valid assumption as long as the change in surface temperature is small [281. In the frequency-domain, the product becomes a convolution, hence Ct 27r Y ChQiw ) 27r w CthQiQo=2 - Ct ei""f i(w - mWS) 3 3ei"wrh(w - mw)6(w - (k + m)w, - w,). (4.10) m=-oo k=-oo Typically, TDTR employs lock-in detection in order to improve signal-to-noise. The lock-in mixes the signal from the detector with a reference signal at w, and with a reference signal with a 900 phase offset, to find in-phase and out-of-phase 97 (quadrature) responses, which filters out all frequencies except for a narrow band' around w 0, which leads to k = -m for all non-zero parts. The complex amplitude of the lock-in response z(T) for a given delay time T is given by [28, 29] z(T) = 4CtQ2O eCikwsh(kws 2 + w,), (4.11) k=-oo where g represents gain in the detection electronics. The measured data in TDTR are the in-phase, x, and quadrature, y, components of the lock-in amplitude as the delay time is varied, z(r) = x(T) + iy(r). 4.1.4 (4.12) Frequency-domain representation of TDTR data As evident in Eq. (4.11), the measured signal in a TDTR experiment is a periodic function of the delay time T, which is represented in the form of a Fourier series. The Fourier series coefficients ak are given by ak = /S7r/w 2 ., [27r /oWZ(T)eiksT - gChQoQiw2 Sh(kw, 47 2 = From Eq. + w0 ). (4.13) (4.13), we can see that the magnitudes of the Fourier coefficients are proportional to h(w) for a discrete set of w values: w = kws + w0 , where k is an integer from -oc to oc. Since the impulse response h(t) is a real function, we can invoke the complex conjugate relation h(w) =*(-w), to find responses h(nws - wO). Thus the frequency responses can be obtained from the Fourier series coefficients as follows, 3 The DC portion of the surface temperature response is rejected by the lock-in, which only detects signal frequency components at the reference frequency. 98 h(nw, + w,) = h(nw. - w,) = 47r 2 gChQ0 Qiw,2 47r 2 4 2 gCthQQw a., 4.-. (4.14) (414 where n is a positive integer. By determining the Fourier coefficients from a measured TDTR response represented as a complex function of the delay time, one is able to determine the frequency response h(w) for w = nw w0 , which is equivalent to determining h(w) from individual FDTR measurements at these frequencies. By Eq. (4.13), determining the Fourier coefficients is possible given a full period of delaytime-domain TDTR data. We refer to this method of extracting frequency response information from TDTR data as "frequency-domain TDTR (fdTDTR)." In some implementations of TDTR, the pump pulses are delayed with respect to the probe [27]. If the delay line is placed after the modulator used to modulate the pump beam at w 0 , an additional phase lag [28] given by exp(iwT) is introduced. In this case, the lock-in response is no longer a periodic function of T. Multiplying the response by exp(-iw0 T) removes the additional phase lag and yields a periodic function described by Eq. (4.11), after which Eq. (4.14) can be used to find frequencydomain responses. A similar problem of finding frequency responses from a lock-in output arises in analyzing acoustic waves measured in a femtosecond pump-probe experiment with a high repetition rate, and an analogous methodology for extracting acoustic frequency responses from the lock-in output has been developed [69, 70]. 4.2 Experimental demonstration of fdTDTR We modify our TDTR setup to enable the collection of a full period of delay time data so that the frequency response analysis described in Eq. (4.14) is possible. We demonstrate the data collection and processing of fdTDTR data using a sample of A1 2 0 3 coated with a thin Al transducer layer (typically -100 99 nm). We go on to show that fdTDTR and FDTR indeed produce the same frequency response information by measuring a sample of A1 2 0 3 coated with a thin Au-Ti transducer layer using the recently completed broadband FDTR system in our laboratory. 4.2.1 Data collection The typical experimental arrangement used in TDTR has been described in numerous works [25, 26, 27, 29]. We introduced our setup [29], which uses a pulsed laser oscillator operating at a center wavelength of 800 nm, with a pulse width of ~200 fs, and a repetition rate of f, = 81 MHz, in Section 1.3.2. An electro-optic modulator (EOM) sinusoidally modulates the pump beam at a frequency f', which we vary from 2 to 12 MHz. To determine the frequency responses of the surface temperature given by Eq. (4.14), we need a full period of delay-time-domain data from T = 0 ns to T 1/fs 12.3 ns. Our existing TDTR setup uses a motorized mechanical delay stage with a 0.5 m travel distance, and passes the probe beam through this delay four times, resulting in a maximum probe delay time of around 7 ns. To obtain an additional 6 ns of delay time, we introduce an additional fixed delay, as illustrated in Fig. 4-4. Thus, the full period of delay time data is collected in two sets. The necessary delay length in the experiment is some amount longer than 1/fe, because measuring the reflectance signal peak at zero delay for both data sets is essential for "stitching" the data sets, as described in Section 4.2.2. To mitigate optical alignment errors and minimize divergence issues, we expand the beam diameter by 4x before the optical delay line, to -8 mm. Accurately extracting high frequency data requires high time resolution at early delay times, where the peak in thermoreflectance occurs. To achieve high time resolution near the thermoreflectance peak, while minimizing data collection time, we vary the speed of our movable delay stage such that it moves slowly for delay times near the peak and more quickly for delay times far away from the peak. Sharp features in the thermoreflectance response only occur near the peak in thermoreflectance, so sub-picosecond resolution is only necessary at these short times. 100 fo additional fixed delay . * red SHG filter ^4 EOM,f. 4x compress pump 400 nm probe 800 nm PBS A/2 expand % filter eND 40 movable delay stage laser, fs 1ox dichroic obj. BS blue detector filter Figure 4-4: TDTR experimental diagram. The additional static delay line combined with the movable delay stage allows for the collection of more than one full period of delay-time-domain data. To further aid the stitching process, the signal-to-noise levels in the two traces are roughly equalized by introducing a neutral density (ND) filter into the path of the probe for the short delay time data set. Due to reflections off of more mirrors, the longer delay time data set has a lower probe power, and hence a lower signal level. The neutral density filter reduces the probe power in the early time trace such that the signal levels in both traces are more matched. Furthermore, the lock-in settings are kept the same for both traces. To demonstrate the fdTDTR data collection and processing, we use a sample that consists of a crystalline substrate of (0001) A1 2 0 3 that was deposited with a 110 nm thick layer of Al using electron beam evaporation. The Al transducer layer thickness was verified with atomic force microscopy by scratching away a small area of Al from 2 the substrate and measuring the step height. The Gaussian pump and probe 1/e radii used were 28 pm and 5 pm respectively. Measurements with larger pump radii, up to 55 pm, produced the same results, but with lower signal-to-noise. The pump radius is chosen to be much larger than the probe radius to mitigate overlap alignment errors. 101 4.2.2 Data stitching The short and long delay data sets are stitched to form the full period of delay time data in Fig. 4-5, which shows both the in-phase, x, and quadrature, y, parts of the complex lock-in amplitude signal as a function of delay time, (see Eq. (4.12)). To combine the delay time data sets, several post-processing steps are used. A peak in the thermoreflectance signal occurs when the pump and probe beams arrive at the sample simultaneously, because at that delay time the probe measures the maximum temperature rise. The long delay time data set is shifted so that the peak in the signal magnitude occurs at T = 1/hf. The short delay time data set is similarly shifted so that the peak occurs atr = 0. Each data set is independently phase-corrected by requiring that the quadrature lock-in signal component, y, does not experience a jump [28 at T = 0 or r This procedure removes any phase that may be added by the electronics. = 1/f,. We find that the necessary phase correction is roughly the same in both data sets to within 1 degree. The data sets are scaled so that the magnitudes at T = 0 and T = 1/f, match. The same scaling factor is used on both x and y to preserve the phase information. This scaling is performed to compensate for the reduction in signal magnitude caused by the added optics in the long delay data set. Even though the signal levels in both data sets are roughly equalized by introducing a neutral density filter into the probe path of the short delay time data set, small differences in the signal magnitudes persist and need to be corrected. After scaling, the data sets are shifted so that the values of x and y just before T = 0 and T = 1/f match. The quality of the data set stitching can be evaluated by observing the delay time region around 6 ns, where the data sets overlap, as shown in Fig. 4-5. Good overlap provides confidence in the data stitching procedure, and in the alignment of the optics during the experiment, demonstrating that the probe does not walk or diverge significantly as the mechanical delay stage is swept. 102 1 0.8 -a, 0.6 X 0.4 0.2 - - -- 0 Im -0.1 CO -0.12 -0.14 -1ek 0 2 4 6 8 10 12 14 - (ns) Figure 4-5: Example of stitching TDTR data sets collected in two parts by introducing an additional fixed delay for measuring long delay time data (> 6 ns). The data sets have been independently phase corrected, shifted and scaled. x and y are the in-phase and quadrature output of the lock-in amplifier respectively, and r is the delay time of the probe with respect to the pump. 4.2.3 Fourier series representation A numerical fast Fourier transform operation on z(T), from T duces Fourier series coefficients values, f nf, f0, ak, = 0 to T = 1/f, pro- which are proportional to h at discrete frequency according to Eq. (4.14). We present our frequency-domain data in terms of the amplitude and phase of the surface temperature frequency response, Re(h) 2 + Im(h) 2 and # = tan- 1 (Im(A)/Re(h)). R has a relative mag- nitude with arbitrary units, while # is the absolute phase with units of angle. 0(f) where R = and R(f) can be directly compared to an equivalent FDTR measurement. Figure 4-6(a) presents R and 4 obtained from the lock-in output data shown in Fig. 4-6(b), which were collected at three modulation frequencies: 4, 8 and 12 MHz. A continuous frequency dependence could be obtained [69 if 103 f, could be varied up to f,/2. Signal-to-noise issues with our existing TDTR system limit our maximum pump modulation frequency to -12 Fig. 4-6(a) have frequency gaps. MHz, so the frequency-domain data curves in However, since thermal responses typically lack sharp resonant features, filling in the gaps is not crucial for practical purposes. Representing TDTR data in the frequency-domain leads more readily to a physical interpretation than a conventional delay-time-domain representation such as that of Fig. 4-6(b). In the delay-time-domain, each different pump modulation frequency results in a different curve which must be evaluated separately or using global fitting strategies, while in the frequency-domain, all data from different collapse into a single curve. f, measurements The amplitude data depend on the amplitude factor that generally varies as we change f, because of the variations in the modulation efficiency of the EOM and the sensitivity of the detection electronics, as well as drift of the laser energy. The delay-time-domain signal magnitude jump at T =0 should be independent of the pump modulation frequency. We normalize each delay-timedomain curve to the magnitude jump at T = 0, using the same normalization factor on both x and y to preserve absolute phase information. Thus, amplitude data from multiple f, measurements form a single curve seen in the top panel of Fig. 4-6(a). The phase data shown in the bottom panel form a single curve without any calibration effort. 4.2.4 Frequency upper limit The high time resolution inherent in TDTR (typically limited by the pulse width) enables the extraction of very high frequency data components. However, at high frequencies, the Fourier coefficients are reduced, resulting in a poorer signal-to-noise ratio. Analyzing the signal-to-noise ratio in our frequency-domain data is challenging. As a first approximation, we estimate the noise in our delay-time-domain data by smoothing the measured data using a moving average filter as shown in Fig. 4-7(a). The smoothed data curve is subtracted from the raw data to obtain the noise data shown in Fig. 4-7(b). The frequency spectrum of the noise data in Fig. 4-7(b) is found in the same way 104 (a) 100 o 4MHz S8 MHz o 12 MHz Ca CU -1 -30 -50 17 -70 -9 ('~1. - 0 200 400 600 f (MHz) 800 1000 (b) ------ 4 MHz -- 8 MHz 12 MHz 1 x 0.5 0 U ................. 12 MHz -0.121-O - -0. 0 2 6 -r (ns) 4 8 10 12 Figure 4-6: Room temperature TDTR data for a sample of A12 0 3 with an Al transducer layer, represented in (a) the frequency-domain and (b) the delay-time-domain. Data are shown for pump modulation frequencies of 4, 8 and 12 MHz. 105 (a) -raw data ---smoothed data. 1 x 0.51 0 -0.1 -0.12 cd -0.14 110 10-2 10 -1 - (ns) 100 101 101 -1 100 101 0.021 (b) 0.01 x 0 -0.01' 0.0 0.00 2- 0 -0.00 2- -00.nn 10 - - 10 -r (ns) Figure 4-7: A noise estimate is obtained by smoothing the raw delay-time-domain data, as shown in (a). Subtracting the smoothed curve from the raw data produces the noise data shown in (b). The delay-time-domain data shown here was collected with f0 = 8 MHz on the Al-coated A1 2 0 3 sample discussed above. 106 as the frequency spectrum of the raw delay-time-domain data in Fig. 4-7(a), and the two spectra are compared in Fig. 4-8. For frequencies below 1 GHz, the data amplitude is >2 orders of magnitude larger than the noise amplitude. At 10 GHz, the data amplitude is ~1 order of magnitude larger than the noise amplitude. 100 10-1 2 - --- c -E a -2b 10 10 -3 10 noise spectrum 10-7 0 _0- data spectrum 10 10 10 10 10 10 101 f (Hz) Figure 4-8: Comparison of the frequency response amplitude obtained from the raw data in Fig. 4-7(a) and the noise data in Fig. 4-7(b). Determining the true upper limit of our frequency-domain data will require further modeling work. response regime. One of the key assumptions of our analysis was that of a linear At early times (<5 ps) following short pulse heating in metals, electrons are excited to high energy levels and have temperatures of several thousand degrees, resulting a highly non-linear regime. Early time data has the most influence on high frequency data, but will have some small influence on lower frequency data as well. As a conservative upper limit estimate, we report our frequency-domain data out to 1 GHz. 4.2.5 Direct comparison of fdTDTR and FDTR data Using a recently completed broadband FDTR system in our lab, based on the design of Ref. [18], we make a direct comparison of FDTR and fdTDTR data. The sample 107 studied was a substrate of crystalline (0001) A1 2 0 3 , coated by electron beam evaporation with 5 nm of Ti followed by 160 nm of Au. Ti was used as a stiction layer to improve the adhesion and uniformity of the Au layer. Au was selected as a transducer material due to signal-to-noise issues with our FDTR system. -The high thermoreflectance response of Au at our FDTR probe laser wavelength of 532 nm results in a good signal-to-noise ratio. Testing on samples with Al transducer layers with our FDTR system resulted in a low thermoreflectance response, and correspondingly a low signal-to-noise ratio. Another signal-to-noise constraint with our FDTR system is the need for small pump and probe diameters, typically <3 pm, to achieve a high enough pump fluence. The peak power output of our FDTR pump laser is ~150 mW. Given the spot size constraints on our FDTR system, to directly compare data sets we needed to adapt our TDTR system to be able to collect data at comparably small pump and probe diameters. Replacing our 10x microscope objective shown in Fig. 4-4 with a 50x microscope objective, along with careful alignment of all the optics, was sufficient to achieve Gaussian 1/e2 TDTR pump and probe diameters of 3 pm and 2 pm respectively. The FDTR data was collected with equal pump and probe diameters of 2 pm. Figure 4-9 shows a comparison of fdTDTR and FDTR data collected on the same sample of A1 2 0 3 with a Ti-Au transducer layer. We find reasonable agreement between the <(f) data sets in the overlapping frequency range. 4.3 Thermal model analysis of fdTDTR data In Section 1.3.3, we outlined a solution to the thermal diffusion equation that is widely used for analyzing TDTR and FDTR responses [28, 29, 31]. Our model outputs the frequency-domain response of the surface temperature, h, which we compare to our measured frequency-domain data. We fit our data simultaneously for the substrate thermal conductivity, kaub, and the thermal interface conductance between the substrate and the transducer film, G, using a least squares fitting routine. All 108 0 c -20- FDTR o fdTDTR 0 0 0 -40 -60 -80 0 -100 0 2 10 10 10 3 10 f (MHz) Figure 4-9: Comparison of phase data from fdTDTR and FDTR measurements on the same sample as described in the text. FDTR data courtesy of Samuel Huberman. other material parameters are set to literature values. Either the relative amplitude, R(f), or absolute phase, #(f), data sets may be used for fitting. We begin by examining the fdTDTR data we obtained for a crystalline substrate of (0001) A12 03 that was deposited with a 110 nm layer of Al. The top part of the Al film is modeled as an isothermal layer to mimic energy deposition into a finite depth. The isothermal layer is modeled as having no radial thermal conductivity and a high cross-plane thermal conductivity. We begin by choosing an isothermal layer thickness of 10 nm, as was done in Ref. [28], which is comparable to the optical skin depth in Al. By fitting the data up to 200 MHz, as shown by the dotted lines in Fig. 4-10, we find best fit values of kAIos3 33.7 W/mK and G = 100 MW/m 2 K from R(f), and kA 2 o 3 = 38.6 W/mK and G 105 MW/m2 K from 4(f). The literature value [71] for kAl 2o3 in the (0001) direction is 41.7 W/mK. In spite of recovering close to the literature value of A1 2 0 3 thermal conductivity, the model fails to capture the phase behavior at high frequencies, suggesting the need to properly model the transport processes in the Al. Fast non-equilibrium electronic diffusion during -1 ps following short-pulse excitation deposits the pump energy over 109 a much larger depth than the optical skin depth of ~7 nm [72]. For the purposes of this work, with the main focus on the experimental methodology rather than on non-equilibrium dynamics in a metal following a femtosecond excitation [67, 68], we proceed by finding the isothermal layer thickness that most closely matches our phase data at high frequencies. We find that an isothermal layer thickness of 25 nm produces a good fit to our high frequency data, as shown by the solid lines in Fig. 4-10. A fit with a 25 nm isothermal layer yields best fit values of kAl 2 o 3 = 34.6 W/mK and G = 103 MW/m2 K from R(f), and kAl 2 o 3 = 40 W/mK and G = 110 MW/m 2 K from 0(f). The choice of isothermal layer thickness is important for modeling the high frequency responses, but yields nearly the same thermal conductivity values, because low frequency data are more sensitive to the substrate thermal conductivity, as will be shown in Section 4.3.1. The values of kAl 2 o 3 and G obtained by fitting data in the frequency-domain representation also provide good fits to delay-time-domain data. The influence of how the metal layer is modeled is apparent in the high frequency phase data. As shown in Fig. 4-11(a), the effect of reducing the isothermal layer thickness to 10 nm or increasing it to 40 nm is negligible below 100 MHz, but becomes increasingly important at high frequencies. In fact, isothermal layers of 10 and 40 nm thicknesses fail to describe the high frequency data even if we allow both kA1 2 o 3 and G to vary, as shown in Fig. 4-11(b). Admittedly, accounting for non-equilibrium electronic diffusion in Al with an isothermal layer is a crude approximation. More accurate modeling of the heat transport in the metal transducer layer, for example with a two-temperature model [73], would be the next logical step for improving the accuracy of modeling high frequency responses. The importance of such analysis in interpreting high-frequency FDTR responses in terms of possible non-diffusive effects has been pointed out in a concurrent study [74]. 4.3.1 Sensitivity analysis We can quantify the thermal model sensitivity to a particular parameter, 13, such as the substrate thermal conductivity or the interface conductance, by considering the logarithmic derivative of the model's response with respect to that parameter [75], 110 10 c 10 10-2 0 -6-70- -900 10 10 10 2 10 3 f (MHz) Figure 4-10: Thermal model best fits from simultaneously varying the A12 0 3 thermal conductivity and the Al-A1 2 0 3 thermal interface conductance are shown, assuming either a 10 nm (dotted lines) or a 25 nm (solid lines) isothermal Al layer. Frequencydomain surface temperature amplitude, R, and phase, <5, response derived from room temperature TDTR measurements (open symbols). 111 -30 (a) -50- -70 -90 --30 (b) -50 CD -70 ..... ........ ............. -90 0 200 600 400 f (MHz) 800 1000 Figure 4-11: <(f) data and model curves assuming isothermal Al layer thicknesses of 25 nm (solid lines), 10 nm (dotted lines) and 40 nm (dashed lines). (a) Curves derived from only varying the isothermal layer thickness, holding all other model parameters constant, and (b) best fit model curves allowing both kA 2 o 3 and G to vary. 112 S-dln R = d ln # S d#-ch = ' .n d In 3 (4.15) Sensitivity curves for our Al coated A12 0 3 sample are plotted in Fig. 4-12, showing the thermal model sensitivity to kAl 2 0 3 , G, and the isothermal Al layer thickness, di,,. Figure 4-12 shows that the thermal model is reasonably sensitive to kAI20 3 out to -20 MHz in R and -200 MHz in #, and at high frequencies, the thermal interface conductance has a more dominant contribution to the signal than the A12 0 3 thermal conductivity. 0.8 0.6 kAl o 0.4 ISO d 2 3 / CU -G C 0.2 C )I 201 ,-. ~%.. / 15F 101 -6- NN - 5 N ------------ - 100 - - - . N. N . 101 *~~--~ - - U) 102 103 104 f (MHz) Figure 4-12: Thermal model sensitivity plots as per Eq. (4.15), using G = 110 MW/m 2 K, kAl 2 o 3 = 41.7 W/mK, an isothermal Al layer thickness of di , = 25 nm, a non-isothermal Al thickness of 85 nm with kAl = 237 W/mK, and literature values of volumetric specific heats. In addition to model sensitivity, we can evaluate the quality of the model fit by considering the summed squares of the residuals between the model and the data, x 2 , for a range of kAl 2 0 3 and G values. Figure 4-13 plots variations in the best fit values of kA1203 and G, where 113 (X 2 _ / xmin is the best fit for 10% 2 value. mu~J1 The x 2 contours indicate that R(f) data is more sensitive to G, while #(f) data is more sensitive to kAl 2 3- This explains why the phase data yield a significantly better - accuracy in measuring kAl 2 o 3 (a) 110 cm E 105 100 0.5 95 kA1 2 o3 38 36 34 32 (b) (W/mK) 120 2 115 0.5 0.1 E 110 0.2 0.01 105 100 36 38 kA1203 40 (W/mK) 42 44 i given a range of kAl 2 o3 and G values varied Figure 4-13: Contours of (x2 _ up to 10% about the best fit values. Our model uses an isothermal Al layer thickness of 25 nm and includes data up to 1 GHz for (a) the frequency response amplitude R(f) and (b) the phase of the frequency response 0(f). 114 4.4 Comparison of fdTDTR data for various samples To further test our fdTDTR method, and to evaluate the utility of using this technique for phonon mean free path spectroscopy, we studied three substrate materials (fused silica, (0001) A1 2 0 3 , and (100) Si) spanning a range of thermal diffusivities, and two different transducer materials (Al and Ti-Au). There have been conflicting reports in the literature regarding the observation of frequency-dependent substrate thermal conductivity based on the transducer layer material used. A BB-FDTR study found a frequency dependent behavior for a Si substrate with a 50 nm thick Au transducer with a 5 nm Cr stiction layer [18], while another FDTR study found no frequency dependence using an 80 nm thick Al transducer [74]. Our measurements with Al transducers do not indicate non-diffusive behavior, in agreement with the observations of Ref. [741. We tested both 110 nm thick (see Fig. 4-14) and 60 nm thick (see Fig. 4-15) Al layers. The thermal model curves shown in Figs. 4-14 to 4-17 assume bulk values for the substrate thermal conductivities, where the only fitting parameter was the thermal interface conductance between the metal transducer and the substrate. In addition to Al transducers, we also investigated Au-Ti transducers. Ti was used as a stiction layer to promote adhesion of the Au. Measuring fdTDTR data on Au transducers is particularly challenging given our pump and probe wavelengths. A sharp hot electron peak is apparent in the first 2 ps of TDTR data, as shown in Figs. 4-18(c) and 4-18(d). This sharp peak complicates the phase correction and stitching procedures that we discussed in Section 4.2.2, necessitating extra care during data collection and processing. It is important to scan the delay stage slowly enough to fully resolve the electronic peak, and to keep the signal-to-noise levels in both the long and shot delay time data traces nearly equivalent. For a thick Au transducer that consisted of 160 nm of Au with a 5 nm Ti stiction layer, we observe comparable non-diffusive behavior to that observed for Al transducers (see Fig. 4-16). We find that using a 80 nm thick isothermal surface layer in 115 the Au produces reasonable fits to our high frequency data if we assume bulk values of substrate thermal conductivity. Measurements on a thiner Au transducer (55 nm of Al with 5 nm of Ti) similar to that of Ref. [18], however, produced poor fits, even assuming a 55 nm thick isothermal surface layer (see Fig. 4-17). The thickness of the Au in this case was thinner than the hot electron diffusion length in Au, which has been reported to be 100 nm [76]. The Ti stiction layer was likely heated by these non-equilibrium hot electrons before they deposited their energy to the Au lattice. This supposition is supported by the shape of the early delay time TDTR data shown in Fig. 4-18. We observe a dip followed by a rise in the in the TDTR amplitude signal at early delay times for the 55 nm Au with a 5 nm Ti transducer layer (see the inset of Fig. 4-18(d)), but not for the 160 nm Au with a 5 nm Ti transducer layer (see the inset of Fig. 4-18(c)). Such a signal (see inset of Fig. 4-18(d)) could arise from the excitation of phonons in the Ti layer by hot electrons in the Au [68]. Au has weak electron-phonon coupling, whereby it takes a longer time for hot electrons to relax to phonons compared to other metals like Ti which have stronger electron-phonon coupling. Thus, the hot electrons in the Au have time to excite electrons and phonons in the Ti layer before relaxing to phonons in the Au. The hot Ti phonons will also conduct heat to the Au layer. Thus, the temperature of the Au phonons increases before decreasing again as heat conducts into the underlying substrate. This behavior needs to be accounted for in thermal modeling of TDTR and FDTR data that uses multilayer metallic transducers where one of the metals has weak electron-phonon coupling [68]. By not including this effect in our thermal model, we observe poor agreement with our fdTDTR data (see Fig. 4-17). A thicker Au layer like that in Fig. 4-16 seems to mitigate this effect, allowing for hot Au electrons to relax by interacting with Au phonons before heat conducts to the Ti and into the substrate. Figure 4-19 shows fdTDTR data collected for a polycrystalline Bi 2Te 3 sample coated in a 90 nm Al transducer layer using a pump beam 1/e 2 diameter of 55 nm and a probe diameter of 10 pm. Thermal model fits using isothermal layer thicknesses of 25 nm and 10 nm are also shown. With di, = 10 nm, the model poorly fits the phase data, but with di, = 25 nm the model fits well and produces a reasonable 116 100 . e -1 C 110 nm Al transducer ofused silica 1-2 SAl O 2 3 o Si 10-3 -20 -40 U> -V -60 ~.. ~ -801 -100' 0 1i 101 10 10 f (MHz) Figure 4-14: fdTDTR data from Si, A1 2 0 3 , and fused silica substrates with a 110 nm thick Al transducer. Lines show thermal model assuming bulk properties and using dj = 25 nm. bulk thermal conductivity of kBi2 Te3 = 1.32 W/mK. Previous TDTR measurements on semiconductor alloys have reported a thermal conductivity dependence on pump modulation frequency, which was attributed to non-diffusive transport in the alloy [77]. Recent theoretical and experimental work has suggested that observations of an apparent frequency dependence could result from an anisotropic failure of the Fourier law, improper modeling of electron-phonon coupling in the metal transducer, or from 117 100 101 0~ It 0-2 60 nm Al transducer 03 0 fused silica Al 2 0 3 Si 10-3 -20 -40 0D -e- -60 -80 0- -100'100 101 102 f (MHz) 103 Figure 4-15: fdTDTR data from Si, A1 2 0 3 , and fused silica substrates with a 60 nm thick Al transducer. Lines show thermal model assuming bulk properties and using dis, = 25 nm. nonequilibrium transport near the interface that renders the usual radiative boundary condition inadequate [73, 74]. We find that accounting for a finite skin depth of optical penetration as well as some length of electron superdiffusion with the crude model of an isothermal layer at the top surface of the metal transducer produces a reasonable fit to our fdTDTR data with bulk material properties for all the substrates we have examined. Our findings support recent works that advocate for improved modeling of nonequilibrium electron and phonon transport in the analysis of FDTR and TDTR 118 100 o. a. ~-1 10 10- - 160 nm Au, 5 nm Ti transducer fused silica SAl O 2 3 10 310Si -40 - ... ...... o ) -20 -o -60 ., -1000 10 - -80. 10 10 2 10 3 f (MHz) Figure 4-16: fdTDTR data from Si, A1 2 0 3 , and fused silica substrates with a 160 nm Au transducer with a 5 nm Ti stiction layer. Lines show thermal model assuming bulk properties and using diso = 80 nm. data [73, 74]. 4.5 Summary and future directions The frequency-domain representation helps uncover aspects of the measurement physics which remain obscured in a traditional TDTR measurement, such as the importance of modeling the details of the heat transport in the metal transducer film for analyzing 119 100 -1 -Z 10-2 55 nm Au, 5 nm Ti transducer 13 fused silica o A 2O 3 10- -20 Al -40- 0- 0 1000 10 10 10 2 10 3 f (MHz) Figure 4-17: fdTDTR data from Si, A12 0 3 , and fused silica substrates with a 55 nm Au transducer with a 5 nm Ti stiction layer. Lines show thermal model assuming bulk properties and treating the entire Au layer as isothermal. high frequency responses. We have detailed a modified TDTR technique that allows for transforming TDTR data collected in the delay-time-domain into the frequency-domain, a representation equivalent to that of FDTR techniques. A single TDTR measurement provides the same information as sweeping through many different modulation frequencies in FDTR. The high time resolution inherent in TDTR measurements enables the extraction of very high frequency content, up to 1 GHz or more, which goes well beyond 120 (b) (a) 100 100 10, - 0 0 1 0 Ce Ce 2 3 x (ns) 4 0 100 50 -r (ps) 5 0 6 2 3 t(ns) 100 50 (ps) C4-C 5 4 6 (d) (c) 100 100 100 100 0 100 5 x 0 1 2 3 -r (ns) 4 5 50 (ps) (ps) 6 100 0 1 2 3 -r(ns) 4 5 6 Figure 4-18: Comparison of early delay time TDTR amplitude signal shapes for different transducer layers on substrates of A12 0 3 including (a) 110 nm Al, (b) 60 nm Al, (c) 160 nm Au with 5 nm Ti, and (d) 55 nm Au with 5 nm Ti. Insets zoom in on the peak near T = 0 where the pump and probe pulses arrive at the sample surface simultaneously. the current capabilities of FDTR techniques [18, 24]. The method only requires a small modification of a conventional TDTR experiment, i.e., the extension of the optical delay range up to a full repetition rate period, and can be easily implemented in any laboratory possessing a standard femtosecond pump-probe apparatus with a high repetition rate. The frequency-domain representation has revealed that while the standard heat diffusion equation model works well at frequencies below ~200 MHz, higher frequency responses are affected by electron superdiffusion in the metal transducer film. This effect will be even more pronounced for metals with weaker electron-phonon coupling such as gold [741. The described methodology not only al- lows a direct comparison of TDTR and FDTR data and yields frequency responses 121 10 U 10 10 10 -40 -60-80 -100 100 10 10 72 103 f (MHz) Figure 4-19: fdTDTR data (open symbols) from a polycrystalline Bi 2Te 3 sample with a 90 nmr Al transducer layer. Model fits shown for di,; = 25 nm (solid lines) and di,,.= 10 nm (dashed lines). #(f) fit with dir8 = 25 nm gives kBi 2Te 3 = 1.32 W/mK and G = 17 MW/m 2 K. at hitherto unattainable high frequencies, but also provides a physically intuitive way of analyzing TDTR measurements. 122 Chapter 5 Summary and Outlook In this thesis we explore both spatially periodic and time-harmonic excitations for measuring non-diffusive conduction heat transport. We utilize an indirect method for determining which phonons are important for transporting heat by measuring thermal conductivity as a function of thermal length scale keff(L) in quasiballistic transport regimes. Measurements of keff(L) can be linked to thermal conductivity accumulation functions through modeling with heat flux suppression functions. We numerically solved the full spectral Boltzmann transport equation (BTE) for transient thermal grating experiments on substrates of Si and PbSe at 300 K using phonon dispersion relations and lifetimes for all six phonon branches from density functional theory calculations. We also solved the gray-medium problem analytically. Our simulations reveled that an approximate model based on summing gray-medium solutions could reasonably model the behavior in PbSe, and that the two-fluid model from Ref. [8] effectively describes Si [78]. We went on to solve the inverse problem of reconstructing thermal conductivity accumulation functions from measured effective thermal conductivities and modeled heat flux suppression functions. We found that a suppression function derived from the gray-medium BTE could reasonably reconstruct the thermal conductivity accumulation function for PbSe, and even that of Si to some extent [78]. We also demonstrated that a Fuchs-Sondheimer suppression function could be used to reconstruct the thermal conductivity accumulation function of bulk Si from TTG measurements on Si membranes of various thicknesses [54]. 123 Finally we explored the viability of deriving heat flux suppression functions experimentally, using the Si membrane data as an ideal test case and found that such an approach is complicated by the sharp features in the differential thermal conductivity function. To measure conduction over length scales of 100 nm, we explored a technique that entails fabricating a ID wire grid polarizer on the surface of a sample of interest. The 1D metal grating acts to minimize light transmission into the underlying substrate, so that primarily only the metal wires are heated during a time-domain thermoreflectance (TDTR) measurement with linearly polarized pump and probe beams aligned perpendicular to the grating transmission axis. We developed a set of design criteria and performed electromagnetic wave finite element simulations to design a suitable grating. We then fabricated gratings on substrates of Si, fused silica and polycrystalline Bi 2Te 3 . Transmission measurements on the transparent fused silica substrate showed that the gratings performed well as linear polarizers for our 800 nm probe and 400 nm pump beams. TDTR measurements indicated non-diffusive transport in Si, and diffusive transport in fused silica and polycrystalline Bi 2Te 3 , indicating that the heat carrying phonons in polycrystalline Bi2 Te 3 have mean free paths (MFPs) of less than 100 nm. We also identified that even small amounts of transmitted pump and probe light that directly excites electron-hole pairs in the substrate and probes the substrate's reflectance response can have an appreciable effect on the data, adding to experimental uncertainty. An alternate approach that eliminates any direct optical excitation of the substrate would be preferable. To study length scales smaller than 100 nm without the need for microfabrication, we developed a method for extracting high frequency response information from TDTR data. Our approach allows TDTR data to be represented in a form equivalent to frequency-domain thermoreflectance (FDTR) data [791. At high excitation frequencies, which correspond to shallow thermal penetration depths, FDTR results including frequencies up to 200 MHz have been reported to exhibit non-diffusive behavior that enables the reconstruction of thermal conductivity accumulation functions from experimental measurements [18]. To date, our TDTR measurements, which allow for frequencies of 1 GHz, have not exhibited deviation from the Fourier heat equation. 124 We suspect that prior observations did not account for electron superdiffusion in the metal transducer, which could explain the observations of apparent non-diffusive behavior in the substrate. Nevertheless, our method of analyzing TDTR data allows for a direct comparison of FDTR and TDTR data, and provides a more physically intuitive representation of TDTR data in the form of the frequency response. Future modeling work may help reveal regimes where information about non-diffusive transport in the substrate can be extracted from high frequency response data. The high frequency data could also be useful for studying electron transport in the metal transducer and transport across the transducer-substrate interface. Future work to advance phonon mean free path spectroscopy should focus on a few key areas. Methods for interpreting experimental data should be reevaluated with regards to the applicability utilizing the heat equation when non-equilibrium effects like hot electron superdiffusion and non-diffusive phonon transport are present. The indirect approach of determining thermal conductivity accumulation functions from measurements of length scale dependent thermal conductivity using heat flux suppression functions [51] has not been rigorously proven, although applications of this method have produced good agreement with first principles calculations [51, 78, 80, 81]. Defining heat flux suppression functions can prove difficult, particularly for experimental length scales that depend on thermal conductivity, such as the thermal penetration depth in FDTR measurements. Future experimental efforts should not only focus on generating and measuring small thermal length scales, but should also emphasize how amenable the measurement geometry is to extracting phonon spectral information as unambiguously as possible on the basis of present knowledge. For example, theoretical analysis of the thin membrane geometry with diffusely scattering boundaries is rigorous [54], and the thermal length scale (the membrane thickness) is well defined. In contrast, experiments with multiple thermal length scales [58] or length scales that depend on thermal diffusivity [18] generate more uncertainty in interpretation. Additionally, improper modeling of interface transport can lead to observations of apparent, but not actual, non-diffusive effects [74], motivating the need for further studies of transport near interfaces and physical boundaries. 125 126 Bibliography [1] G. Chen. Nonlocal and nonequilibrium heat conduction in the vicinity of nanoparticles. J. Heat Transfer, 118(3):539-545, 08 1996. [2] G. Chen. Nanoscale Energy Transport and Conversion. Oxford University Press, Inc., 2005. [3] M. E. Siemens, Q. Li, R. Yang, K. A. Nelson, E. H. Anderson, M. M. Murnane, and H. C. Kapteyn. Quasi-ballistic thermal transport from nanoscale interfaces observed using ultrafast coherent soft x-ray beams. Nat. Mater., 9(1):26-30, 01 2010. [4] A. J. Minnich, J. A. Johnson, A. J. Schmidt, K. Esfarjani, M. S. Dresselhaus, K. A. Nelson, and G. Chen. Thermal conductivity spectroscopy technique to measure phonon mean free paths. Phys. Rev. Lett., 107:095901, Aug. 2011. [5] A. J. Minnich. Towards a microscopic understanding of phonon heat conduction, 2014. arXiv:1405.0532, submitted. [6] T. Klitsner, J. VanCleve, H. Fischer, and R. Pohl. Phonon radiative heat transfer and surface scattering. Phys. Rev. B, 38:7576-7594, Oct. 1988. [7] J. Schleeh, J. Mateos, I. Iniguez-de-la Torre, N. Wadefalk, P. A. Nilsson, J. Grahn, and A. J. Minnich. Phonon black-body radiation limit for heat dissipation in electronics. Nat. Mater., advance online publication, 11 2014. [8] A. A. Maznev, J. A. Johnson, and K. A. Nelson. Onset of nondiffusive phonon transport in transient thermal grating decay. Phys. Rev. B, 84:195206, Nov. 2011. [9] A. Majumdar. Microscale heat conduction in dielectric thin films. J. Heat Trans- fer, 115(1):7-16, 1993. [10] C. Dames and G. Chen. Thermal conductivity of nanostructured thermoelectric materials. In D. M. Rowe, editor, Thermoelectrics Handbook: Macro to Nano, chapter 42. CRC Press, 2006. [11] F. Yang and C. Dames. Mean free path spectra as a tool to understand thermal conductivity in bulk and nanostructures. Phys. Rev. B, 87:035437, Jan. 2013. 127 [12] Z. Tian, S. Lee, and G. Chen. Heat transfer in thermoelectric materials and devices. J. Heat Transfer, 135(6):061605-061605, 05 2013. [13] 0. Delaire, J. Ma, K. Marty, A. F. May, M. A. McGuire, M-H. Du, D. J. Singh, A. Podlesnyak, G. Ehlers, M. D. Lumsden, and B. C. Sales. Giant anharmonic phonon scattering in pbte. Nat. Mater., 10(8):614-619, 08 2011. [14] E. Burkel. Inelastic Scattering: of X-Rays with Very High Energy Resolution (Springer Tracts in Modern Physics). Springer, 1991. [15] A. A. Maznev, K. J. Manke, K.-H. Lin, K. A. Nelson, C.-K. Sun, and J.-I Chyi. Broadband terahertz ultrasonic transducer based on a laser-driven piezoelectric semiconductor superlattice. Ultrasonics, 52(1):1-4, 2012. [16] A. A. Maznev, F. Hofmann, A. Jandl, K. Esfarjani, M. T. Bulsara, E. A. Fitzgerald, G. Chen, and K. A. Nelson. Lifetime of sub-thz coherent acoustic phonons in a gaas-alas superlattice. Applied Physics Letters, 102(4), 2013. [17] B. Daly, K. Kang, Y. Wang, and D. Cahill. Picosecond ultrasonic measurements of attenuation of longitudinal acoustic phonons in silicon. Phys. Rev. B, 80:174112, Nov. 2009. [18] K. T. Regner, D. P. Sellan, Z. Su, C. H. Amon, A. J. H. McGaughey, and J. A. Malen. Broadband phonon mean free path contributions to thermal conductivity measured using frequency domain thermoreflectance. Nat. Commun., 4(1640), 2013. [19] J. A. Johnson, A. A. Maznev, J. Cuffe, J. K. Eliason, A. J. Minnich, T. Kehoe, C. M. Sotomayor Torres, G. Chen, and K. A. Nelson. Direct measurement of room-temperature nondiffusive thermal transport over micron distances in a silicon membrane. Phys. Rev. Lett., 110:025901, Jan. 2013. [20] A. G. Bell. On the production and reproduction of sound by light. American Journal of Science, (118):305-324, 1880. [21] A. Rosenewaig, J. Opsal, W. L. Smith, and D. L. Willenborg. Detection of thermal waves through optical reflectance. Appl. Phys. Lett., 46(11):1013-1015, 1985. [22] M. Wagner, N. Winkler, and H. D. Geiler. Single-beam thermowave analysis of semiconductors. Applied Surface Science, 50(1A14):373 - 376, 1991. [23] F. Lepoutre, D. Balageas, Ph. Forge, S. Hirschi, J. L. Joulaud, D. Rochais, and F. C. Chen. Micron-scale thermal characterizations of interfaces parallel or perpendicular to the surface. J. Appl. Phys., 78(4):2208-2223, 1995. [24] K. T. Regner, S. Majumdar, and J. A. Malen. Instrumentation of broadband frequency domain thermoreflectance for measuring thermal conductivity accumulation functions. Review of Scientific Instruments, 84(6), 2013. 128 [25] C. A. Paddock and G. L. Eesley. Transient thermoreflectance from thin metal films. J. Appl. Phys., 60(1):285-290, 1986. [26] W. S. Capinski and H. J. Maris. Improved apparatus for picosecond pump-andprobe optical measurements. Review of Scientific Instruments, 67(8):2720-2726, 1996. [27] D. G. Cahill, K. Goodson, and A. Majumdar. Thermometry and thermal transport in micro/nanoscale solid-state devices and structures. J. Heat Transfer, 124(2):223-241, 2002. [28] D. G. Cahill. Analysis of heat flow in layered structures for time-domain thermoreflectance. Review of Scientific Instruments, 75(12):5119-5122, 2004. [29] A. J. Schmidt, X. Chen, and G. Chen. Pulse accumulation, radial heat conduction, and anisotropic thermal conductivity in pump-probe transient thermoreflectance. Review of Scientific Instruments, 79(11):114902, 2008. [30] A. J. Schmidt. Optical characterizationof thermal transportfrom the nanoscale to the macroscale. PhD thesis, Massachusetts Institute of Technology, 2008. [31] C. Xing, C. Jensen, Z. Hua, H. Ban, D. H. Hurley, M. Khafizov, and J. R. Kennedy. Parametric study of the frequency-domain thermoreflectance tech- nique. J. Appl. Phys., 112(10), 2012. [32] H. S. Carslaw and J. C. Jaeger. Conduction of Heat in Solids. Oxford University Press, 2nd edition, 1959. [33] D. Maillet, S. Andre, J. C. Batsale, A. Degiovanni, and C. Moyne. Thermal Quadrupoles: Solving the Heat Equation through Integral Transforms. John Wiley and Sons, LTD., 2000. [34] H. J. Eichler, P. Giinter, and D. W. Pohl. volume 50. Springer-Verlag Berlin, 1986. Laser-induced dynamic gratings, [35] J. A. Rogers, Y. Yang, and K. A. Nelson. Elastic modulus and in-plane thermal diffusivity measurements in thin polyimide films using symmetry-selective realtime impulsive stimulated thermal scattering. Appl. Phys. A, 58(5):523-534, 1994. [36] J. A. Rogers, A. A. Maznev, M. J. Banet, and K. A. Nelson. Optical generation and characterization of acoustic waves in thin films: Fundamentals and applications. Annu. Rev. Mater. Sci., 30(1):117-157, 2000. [37] J. A. Johnson, A. A. Maznev, M. T. Bulsara, E. A. Fitzgerald, T. C. Harman, S. Calawa, C. J. Vineis, G. Turner, and K. A. Nelson. Phase-controlled, heterodyne laser-induced transient grating measurements of thermal transport properties in opaque material. J. Appl. Phys., 111(2):023503, 2012. 129 [38] A. A. Mazncv, K. A. Nelson, and J.A. Rogers. Optical heterodyne detection of laser-induced gratings. Opt. Lett., 23(16):1319-1321, Aug. 1998. [39] A. Ward and D. Broido. Intrinsic phonon relaxation times from first-principles studies of the thermal conductivities of si and ge. Phys. Rev. B, 81:085205, Feb. 2010. [40] R. Peierls. Zur kinetischen theorie der wdrmeleitung in kristallen. Annalen der Physik, 395(8):1055-1101, 1929. [41] A. J. Minnich, G. Chen, S. Mansoor, and B. S. Yilbas. Quasiballistic heat transfer studied using the frequency-dependent boltzmann transport equation. Phys. Rev. B, 84:235207, Dec. 2011. [42] Q. Hao, G. Chen, and M.-S. Jeng. Frequency-dependent monte carlo simulations of phonon transport in two-dimensional porous silicon with aligned pores. J. Appl. Phys., 106(11), 2009. [43] W. H. Press. Numerical Recipes in Fortran 77: The Art of Scientific Computing. Fortran Numerical Recipes. Cambridge University Press, 1992. [44] J. Callaway. Model for lattice thermal conductivity at low temperatures. Phys. Rev., 113:1046-1051, Feb. 1959. [45] M. G. Holland. Phonon scattering in semiconductors from thermal conductivity studies. Phys. Rev., 134:A471-A480, Apr 1964. [46] K. Esfarjani, G. Chen, and H. T. Stokes. Heat transport in silicon from firstprinciples calculations. Phys. Rev. B, 84:085204, Aug. 2011. [47] Z. Tian, J. Garg, K. Esfarjani, T. Shiga, J. Shiomi, and G. Chen. Phonon conduction in pbse, pbte, and pbtei-zse, from first-principles calculations. Phys. Rev. B, 85:184303, May 2012. [48] A. S. Henry and G. Chen. Spectral phonon transport properties of silicon based on molecular dynamics simulations and lattice dynamics. J. Comput. Theor. Nanosci., 5(2):1-12, Feb. 2008. [49] H. Wang, Y. Pei, A. D. LaLonde, and G. J. Snyder. Heavily doped p-type PbSe with high thermoelectric performance: An alternative for PbTe. Advanced Materials, 23(11):1366-1370, 2011. [50] Y. S. Ju and K. E. Goodson. Phonon scattering in silicon films with thickness of order 100 nm. Appl. Phys. Lett., 74(20):3005-3007, 1999. [51] A. J. Minnich. Determining phonon mean free paths from observations of quasiballistic thermal transport. Phys. Rev. Lett., 109:205901, Nov. 2012. [52] M. Grant and S. Boyd. CVX: Matlab software for disciplined convex programming, version 2.0 beta, 2012. 130 [53] M. Grant and S. Boyd. Graph implementations for nonsmooth convex programs. In V. Blondel, S. Boyd, and H. Kimura, editors, Recent Advances in Learning and Control, Lecture Notes in Control and Information Sciences, pages 95-110. Springer-Verlag Limited, 2008. [54] J. C. Cuffe, J. K. Eliason, A. A. Maznev, K. C. Collins, J. A. Johnson, A. Shchepetov, M. Prunnila, J. Ahopelto, C. M. Sotomayor Torres, G. Chen, and K. A. Nelson. Reconstructing phonon mean free path contributions to thermal conductivity using nanoscale membranes. 2014. arXiv:1408.6747, submitted. [55] K. Fuchs. The conductivity of thin metallic films according to the electron theory of metals. Mathematical Proceedings of the Cambridge Philosophical Society, 34:100-108, 1 1938. [56] E. H. Sondheimer. The mean free path of electrons in metals. Advances in Physics, 1:1-42, 1952. [57] A. J. Minnich. Exploring Electron and Phonon Transport at the Nanoscale for Thermoelectric Energy Conversion. PhD thesis, Massachusetts Institute of Technology, 2011. [58] K. M. Hoogeboom-Pot, J. N. Hernandez-Charpak, E. H. Anderson, X. Gu, R. Yang, M. M. Murnane, H. C. Kapteyn, and D. Nardi. A new regime of nanoscale thermal transport: collective diffusion counteracts dissipation ineffi- ciency. 2014. arXiv:1407.0658, submitted. [59] E. Chivez-Angel, J. S. Reparaz, J. Gomis-Bresco, M. R. Wagner, J. Cuffe, B. Graczykowski, A. Shchepetov, H. Jiang, M. Prunnila, J. Ahopelto, F. Alzina, and C. M. Sotomayor Torres. Reduction of the thermal conductivity in free-standing silicon nano-membranes investigated by non-invasive raman ther- mometry. APL Materials, 2(1), 2014. [60] L. Zeng and G. Chen. Disparate quasiballistic heat conduction regimes from periodic heat sources on a substrate. J. Appl. Phys., 116(6), 2014. [61] M. G. Burzo, P. L. Komarov, and P. E. Raad. Minimizing the uncertanties associated with the measurement of thermal properties by the transient thermoreflectance method. IEEE Transactions on Components and Packaging Tech- nologies, 28(1):39-44, Mar. 2005. [62] H. Lloyd. On a new case of interference of the rays of light. The Transactions of the Royal Irish Academy, 17:171-177, Jan. 1831. [63] B. Poudel, Q. Hao, Y. Ma, Y. Lan, A. J. Minnich, B. Yu, X. Yan, D. Wang, A. Muto, D. Vashace, X. Chen, J. Liu, M. S. Dresselhaus, G. Chen, and Z. Ren. High-thermoelectric performance of nanostructured bismuth antimony telluride bulk alloys. Science, 320(5876):634-638, 2008. 131 [64] H.-P. Feng, B. Yu, S. Chen, K. C. Collins, C. He, Z. F. Ren, and G. Chen. Studies on surface preparation and smoothness of nanostructured Bi2 Te 3-based alloys by electrochemical and mechanical methods. ElectrochimicaActa, 56(8):3079 - 3084, 2011. [65] 0. Hellman and D. A. Broido. Phonon thermal transport in Bi 2 Te 3 from first principles. Phys. Rev. B, 90:134309, Oct. 2014. [66] K. Kang, Y. K. Koh, C. Chiritescu, X. Zheng, and D. G. Cahill. Two-tint pumpprobe measurements using a femtosecond laser oscillator and sharp-edged optical filters. Review of Scientific Instruments, 79(11), 2008. [67] S. D. Brorson, J. G. Fujimoto, and E. P. Ippen. Femtosecond electronic heat- transport dynamics in thin gold films. Phys. Rev. Lett., 59:1962-1965, Oct. 1987. [68] G.-M. Choi, R. B. Wilson, and D. G. Cahill. Indirect heating of pt by short-pulse laser irradiation of au in a nanoscale pt/au bilayer. Phys. Rev. B, 89:064307, Feb. 2014. [69] S. Kaneko, M. Tomoda, and 0. Matsuda. A method for the frequency control in time-resolved two-dimensional gigahertz surface acoustic wave imaging. AIP Advances, 4(1), 2014. [70] 0. Matsuda and 0. B. Wright. Time-resolved gigahertz acoustic wave imaging at arbitrary frequencies. IEEE Trans. Ultrason. Ferroelectr. Freq. Control. in press. [71] MTI Corporation. A12 0 3 single crystal. http://www.mtixtl.com/xtlflyers/Al203.pdf. Accessed: 2014-04-01. [72] G. Tas and H. J. Maris. Electron diffusion in metals studied by picosecond ultrasonics. Phys. Rev. B, 49:15046-15054, Jun. 1994. [73] R. B. Wilson, J. P. Feser, G. T. Hohensee, and D. G. Cahill. Two-channel model for nonequilibrium thermal transport in pump-probe experiments. Phys. Rev. B, 88:144305, Oct. 2013. [74] R. B. Wilson and D. G. Cahill. Anisotropic failure of fourier theory in timedomain thermoreflectance experiments. Nat. Commun., 5, Oct. 2014. [75] B. C. Gundrum, D. G. Cahill, and R. S. Averback. Thermal conductance of metal-metal interfaces. Phys. Rev. B, 72(24):245426, Dec. 2005. [76] 0. B. Wright and V. E. Gusev. Ultrafast acoustic phonon generation in gold. Physica B: Condensed Matter, 219-220(0):770-772, 1996. [77] Y. K. Koh and D. G. Cahill. Frequency dependence of the thermal conductivity of semiconductor alloys. Phys. Rev. B, 76:075207, Aug. 2007. 132 [78] K. C. Collins, A. A. Maznev, Z. Tian, K. Esfarjani, K. A. Nelson, and G. Chen. Non-diffusive relaxation of a transient thermal grating analyzed with the boltzmann transport equation. J. Appl. Phys., 114(10), 2013. [79] K. C. Collins, A. A. Maznev, J. Cuffe, K. A. Nelson, and G. Chen. Examining thermal transport through a frequency-domain representation of time-domain thermoreflectance data. Review of Scientific Instruments, 85(12), 2014. [80] D. Ding, X. Chen, and A. J. Minnich. Radial quasiballistic transport in timedomain thermoreflectance studied using monte carlo simulations. Appl. Phys. Lett., 104(14), 2014. [81] H. Zhang, C. Hua, D. Ding, and A. J. Minnich. Length dependent thermal conductivity measurements yield phonon mean free path spectra in nanostructures, 2014. arXiv:1410.6233, submitted. 133