Device Modeling from Atomistic First Principles: theory of the nonequilibrium vertex correction Eric Zhu1, Leo Liu1, Hong Guo1,2 1 Nanoacademic Technologies 2 Dept. Inc. Brossard, QC J4Z 1A7, Canada of Physics, McGill Univ., Montreal, Quebec, H3A 2T8 Canada • Introduction: NEGF-DFT; • 4 critical issues: disorder averaging, band gap, large sizes, verification; • Two examples: localized doping in Si nanoFET; disorder scattering in MRAM; • Summary. Atomic model Continuum model 2012-08-29 NEGF5, Jyvaskyla, Finland page 1 Goal: simulate a transistor from atomic first principles ~10nm ~100nm 1. 2. 3. 4. Doping & disorder, Band gaps, Large sizes, Accuracy. Other physics: phonons, magnons, photons, correlations… Picture from Taur and Ning, Fundamentals of Modern VLSI Devices (10 nm)3 chunk of Si has ~64,000 atoms. 2012-08-29 Current: L=22nm Next: L=16nm NEGF5, Jyvaskyla, Finland DFT: ~1,000 atoms page 2 … and many other systems with different materials This talk: In any real device made of any real material, there is a degree of disorder. ? Such disorder impacts device operation in serious ways. How do we compute these effects from first principles? Can we calculate ? 2012-08-29 NEGF5, Jyvaskyla, Finland page 3 Ex 1: Dopant fluctuation gives rise to device-to-device variability Huge device to device variability. If every transistor behaves differently, difficult to design a circuit. F.L. Yang et al., in VLSI Technol. Tech. Symp. Dig., pp. 208, June 2007. 2012-08-29 NEGF5, Jyvaskyla, Finland page 4 Ex 2: roughness scattering increases resistance of Cu interconnects With Daniel Gall of RPI. $: SRC 2012-08-29 NEGF5, Jyvaskyla, Finland page 5 Ex. 3: disorder effect in topological insulator Bi2Se3 Conductance: Experiment (Hasan etal) Ab initio (Zhao etal) Calculated spin direction Top surface Zhao, H.G. etal Nano Lett. 11, 2088 (2011). Wang, Hu, H.G. PRB 85, 241402 (2012) 2012-08-29 Bottom surface NEGF5, Jyvaskyla, Finland page 6 Can we calculate realistic device parameters? engineering science New quantum mechanics atomic simulations device modeling Physics materials, chemistry, < 5nm (1000 atoms) physics Semi-empirical device modeling, 10,000 to 100,000 atoms device parameters TCAD Quantitative prediction of quantum transport from atomic first principles without any parameter 2012-08-29 NEGF5, Jyvaskyla, Finland page 7 Method - NEGF-DFT: non-equilibrium density matrix NEGF-DFT ‘DFT’: density functional theory DFT Hˆ (r ) NEGF: Keldysh nonequilibrium Green’s function ~ G dE GR GA dE NEGF DFT NEGF-DFT `DFT’ in NEGF-DFT is not the usual ground state DFT: density matrix of NEGF-DFT is constructed at non-equilibrium. No variational solution. Jeremy Taylor, Hong Guo and Jian Wang, Phys. Rev. B 63, 245407 (2001). M. Brandbyge, J.-L. Mozos, P. Ordejon, J. Taylor, and K. Stokbro, PRB 65, 165401 (2002). 2012-08-29 NEGF5, Jyvaskyla, Finland page 9 1. Within NEGF-DFT: solving the disorder averaging problem Doping and disorder scattering from atomic principles Acknowledgements: Dr. Youqi Ke, Dr. Ke Xia, Dr. Ferdows Zahid, Dr. Eric Zhu, Dr. Lei Liu, Dr. Yibin Hu Drs. Eric Zhu, Leo Liu, and Yibin Hu: development of the NEGF-DFT/CPA-NVC first principles package Nanodsim (nano-device-simulator) – Nanoacademic Technologis Inc. (www.nanoacademic.com). NEGF-DFT/CPA-NVC Youqi Ke, Ke Xia and Hong Guo PRL 100, 166805 (2008); Youqi Ke, Ke Xia and Hong Guo, PRL 105, 236801 (2010); Ferdows Zahid, Youqi Ke, Daniel Gall and Hong Guo, PRB 81, 045406 (2010); Eric Zhu, Lei Liu and Hong Guo, preprint (2012). 2012-08-29 NEGF5, Jyvaskyla, Finland page 10 A tough problem of atomic calculations: disorder scattering Generating many configurations, compute each, and average result very time consuming (Small x, large N) Ax B1 x T. Dejesus, Ph.D thesis, McGill University, 2002. For any theoretical calculation, disorder averaging must be done. How to do it in atomistic calculations at non-equilibrium? 2012-08-29 NEGF5, Jyvaskyla, Finland page 11 To build intuition, let’s solve a toy problem exactly 1 2 1D tight binding chain nearest neighbor coupling on-site energy 2012-08-29 NEGF5, Jyvaskyla, Finland page 12 Self-energies for the leads: L 1 2 R How to handle half-infinite chain ? Self energy: The problem is reduced to 3 sites plus self-energies 2012-08-29 NEGF5, Jyvaskyla, Finland page 13 Physical quantities and NEGF Express physical quantities in terms of NEGF: average over disorder configurations L R The problem is reduced to calculate disorder averaged NEGF 2012-08-29 NEGF5, Jyvaskyla, Finland page 14 L 1 2 R Disorder average can be done exactly for the 3-site model 2012-08-29 NEGF5, Jyvaskyla, Finland page 15 Exact solution of the 3-site toy model: In general, the number of configuration is 2N (N is the number of disorder sites). It is impossible to enumerate and compute all configurations for large N. We need a better “statistical approach” 2012-08-29 NEGF5, Jyvaskyla, Finland Coherent potential Approx. page 16 CPA - well established formalism Ax B1 x When there are impurities, translational symmetry is broken. Coherent Potential Approximation (CPA) is an effective medium theory that averages over the disorder and restores the translational symmetry. So, an atomic site has x% chance to be occupied by A, and (1-x)% chance by B. P. Soven, Phys. Rev. 156, 809 (1967). T ( E, V ) Tr[G LG R ] r a B. Velicky, Phys. Rev. 183 (1969). Rev. Mod. Phys. 46, 466 (1974) 2012-08-29 NEGF5, Jyvaskyla, Finland page 17 CPA: CPA picture: effective media and are solved from CPA equation Implementation: needs a method that does one atom at a time: LMTO, KKR, etc.. 2012-08-29 NEGF5, Jyvaskyla, Finland page 18 Non-equilibrium density matrix: nonequilibrium vertex NEGF-DFT ~ G dE G R G A dE DFT Average over random disorder: Hˆ (r ) NEGF X ~ G dE G R G A dE T Tr G R l G A r specular part G R G A G R G A diffusive part Take Home message #1: multiple disorder scattering at non-equilibrium is solved by the nonequilibrium vertex correction theory (NVC) and implemented in NEGF-DFT software Nanodsim. Youqi Ke, Ke Xia and Hong Guo PRL 100, 166805 (2008) 2012-08-29 NEGF5, Jyvaskyla, Finland page 19 Essence of Nonequilibrium Vertex Correction (NVC) X Conventional vertex correction, i.e. that appears in computing Kubo formula in disordered metal, is done at equilibrium. NVC is done at non-equilibrium: it is related not only to multiple impurity scattering, but also to the non-equilibrium statistics of the device scattering region. Implementation: LMTO with atomic sphere approximation, plus CPA and NVC, within NEGF-DFT. 2012-08-29 NEGF5, Jyvaskyla, Finland page 20 NVC Equation: some complicated technical details Youqi Ke, Ke Xia and Hong Guo PRL 100, 166805 (2008). 2012-08-29 NEGF5, Jyvaskyla, Finland page 21 Consistency check: CPA-NVC identity NVC CPA CPA The CPA-NVC identity can also be proved analytically at non-equilibrium: CPA and NVC are consistent approximations (Eric Zhu and H.G., 2012). The identity is tested numerically: strong check of the code. 2012-08-29 NEGF5, Jyvaskyla, Finland page 22 NVC solution for the 3-site toy model and 2012-08-29 are solved from NVC equation NEGF5, Jyvaskyla, Finland page 23 Comparison for the 3-site toy model: Excellent ! 2012-08-29 specular part NEGF5, Jyvaskyla, Finland diffusive part page 24 Non-toy system: At equilibrium, fluctuation-dissipation theorem holds. Left hand side has NVC; right hand side does not. This gives a very strict check to the NVC formalism as well as to the numerical implementation. no NVC 2012-08-29 NEGF5, Jyvaskyla, Finland NVC exact page 25 2. Within NEGF-DFT: solving the band gap problem The band gap problem … Acknowledgements: Dr. Youqi Ke, Dr. Wei Ji, Mathieu Cesar, Dr. Eric Zhu, Dr. Lei Liu, Dr. Zetian Mi, Dr. Ferdows Zahid NEGF-DFT-CPA-NVC 2012-08-29 NEGF5, Jyvaskyla, Finland page 26 The band gap problem of local functionals in DFT DFT calculation of band gaps: MBJ computation time is ~LDA 2012-08-29 NEGF5, Jyvaskyla, Finland page 27 Some relevant band gaps for transistors materials: 2012-08-29 NEGF5, Jyvaskyla, Finland page 28 Some relevant effective masses Take home message 2: the band gap problem is practically resolved by MBJ semi-local exchange within LMTO implementation of NEGF-DFT. 2012-08-29 NEGF5, Jyvaskyla, Finland page 29 Experimental data Calculated data MBJ potential + CPA: works. Good agreement with experimental data. 2012-08-29 NEGF5, Jyvaskyla, Finland page 30 3. Within NEGF-DFT: solving the large size problem Solving large problems from self-consistent first principles. Acknowledgements: Dr. Eric Zhu, Dr. Lei Liu (Nanoacademic Technologies Inc.) Dr. Yibin Hu, Mohammed Harb, Vincent Michaud-Roux (McGill) J. Maassan, E. Zhu, V. Michaud-Vioux, M. Harb and H.G., to appear in IEEE Proceedings (2012). 2012-08-29 NEGF5, Jyvaskyla, Finland page 31 Locality: the principle underlying all O(N) methods Locality: the properties of a certain observation region comprising one or a few atoms are only weakly influenced by factors that are spatially far away from this observation region. S. Geodecker Rev. Mod. Phys. (1999) Equilibrium density matrix exhibits decaying property: insulator metal LMTO (nanodsim) LCAO (nanodcal) Example: Si bulk 2012-08-29 NEGF5, Jyvaskyla, Finland page 32 Density matrix computation: DFT – computes potential and energy levels of the device; DFT NEGF – non-equilibrium statistics that fills levels; Hˆ (r ) The self-consistent loop – NEGF-DFT algorithm we use. ~ G dE GR GA dE NEGF Practically, density matrix is divided into two parts: equilibrium and non-equilibrium parts: locality 2012-08-29 NEGF5, Jyvaskyla, Finland no locality page 33 Roadmap for locality-less computation of density matrix of large systems Large: ~20,000 atoms algorithms Do not depend on locality iterative method no preconditioner gmres bicgstab 2012-08-29 qmr direct method with preconditioner Jacobi SOR NEGF5, Jyvaskyla, Finland ILU MCS H-Matrix page 34 Roadmap (cont.) algorithms iterative method nested dissection single wall 2012-08-29 double wall direct method principal layer thin & long NEGF5, Jyvaskyla, Finland pardiso thick & short page 35 Performance: nano-device simulator (nanodsim) Nanodsim has been fully parallelized and optimized for both speed and memory costs. The speed is made nearly O(N) along the transport direction. speed performance memory performance Lx = periodic Ly = 10 nm Lz = 5, 10 ,15, 20 ,25, 30nm 160 cores Benchmark: 160 cores, 3GB / core, 12,800 atomic sites, <30 min/step 2012-08-29 NEGF5, Jyvaskyla, Finland page 36 Solving ~20,000 atomic spheres for open devices at nonequilibrium Open device structure of Si: parallel NEGF-DFT run on 480 cores. Structure Size # of atoms NEGF-DFT run Convergence Leads : 1 X 20 X 1 320 atomic spheres (2880 Orbitals) 1 X 20 X 20 6400 atomic spheres (57600 Orbitals) 107 NEGF-DFT steps, 5.5 min/step, total 9.8 hours Potential 1.0 x 10-5 Charge 1.23 x 10-5 per atom Two probe 2 1 X 20 X 40 12800 atomic spheres (115200 Orbitals) 195 NEGF-DFT steps, 14 min/step, total 46 hours Two probe 3 1 x 20 x 60 19200 atomic spheres (172800 Orbitals) 267 NEGF-DFT steps, 30 min/step, total 134 hours Two probe 1 Potential 1.0 x 10-5 Charge 6.2 x 10-6 per atom Same as above Summary: NEGF-DFT modeling has reached realistic device sizes! 2012-08-29 NEGF5, Jyvaskyla, Finland page 37 If using tight binding model: huge systems can be done Run on a single computing node with 12 cores and 36 GB memory 1,024,000 Si atoms 10.9 nm ×10.9 nm × 173.8 nm time = 2 days • Computation time scales linearly with the channel length • Computation time increases 6~7 times if the cross section doubles • For Lz = 173.8 nm, 1/3 of computation time is spent on surface Green’s function, and 2/3 spent on transmission calculation 2012-08-29 NEGF5, Jyvaskyla, Finland page 38 4. How do we know all are well for real devices ? Bench-marking the NEGF-DFT atomistic model for device simulations Acknowledgements: Lining Zhang (ECE, HKUST), Dr. Ferdows Zahid (Physics, HKU), Dr. Mansun Chan (ECE, HKUST), Dr. Jian Wang (Physics, HKU), Dr. Jesse Maassen (ECE, Purdue), Dr. Eric Zhu (Nanoacademic). NEGF-DFT/CPA-NVC 2012-08-29 NEGF5, Jyvaskyla, Finland page 39 Commercial TCAD tool: 1328 pages of parameter and physics descriptions 2012-08-29 NEGF5, Jyvaskyla, Finland page 40 Hundreds of parameters are needed ! 2012-08-29 NEGF5, Jyvaskyla, Finland page 41 NEGF-DFT/CPA-NVC versus Sentaurus Atomic model: parameter-free Continuum model with external parameters Sentaurus: Drift-diffusion coupled with Poisson solver in real space grids NEGF-DFT/CPA-NVC 2012-08-29 NEGF5, Jyvaskyla, Finland page 42 Potential of Si TFET: p-i-n structure L. Zhang, F. Zahid, M. Chan, J. Wang, H.G. (2012). Sentaurus (green) versus NEGF-DFT (red) Band gaps; doping; disorder; large sizes; computation; … Doping in the channel does not affect the potential profile due to high doping at S/D 2012-08-29 Intrinsic channel 8nm 12nm 14nm T=300K p-i-n tunnel FET (TFET) potential: almost perfect agreement NEGF5, Jyvaskyla, Finland page 43 Verification for MOSFET channels Green: Sentaurus. 2012-08-29 NEGF5, Jyvaskyla, Finland Red: Nanodsim page 44 New: non-uniform doping – delta doping Atomistic treatment of doping (P-doped) within CPA formalism Red – NEGF-DFT Green – Sentaurus with Fermi statistics Black – Sentaurus with Boltzman statistics 2012-08-29 NEGF5, Jyvaskyla, Finland page 45 Full double gate FET simulation: LG gate Tox LS oxide LD TSi n+ p n+ Tox sS oxide sD gate 2012-08-29 NEGF5, Jyvaskyla, Finland page 46 I-V characteristics I-V characteristics calculated by atomic model are in good agreement with NanoMos (effective mass model). Atomic model can go much further: surface roughness scattering, inhomogeneous doping, new materials, etc. Nanodsim (self-consistent atomic) 2012-08-29 NEGF5, Jyvaskyla, Finland NanoMos (Zhibin Ren’s thesis) page 47 Example 1: localized doping NEGF-DFT-CPA-NVC Localized doping suppresses off-state source-to-drain tunneling and reduces performance variability. Acknowledgement: Dr. Jesse Maassan (ECE, Purdue) Jesse Maassan & H.G. preprint (2012). 2012-08-29 NEGF5, Jyvaskyla, Finland page 48 New idea: suppressing S-to-D off-state tunneling 2012-08-29 NEGF5, Jyvaskyla, Finland page 49 Example 2: MRAM simulations Increasing spin transfer torque (STT) by impurity doping. Acknowledgements: Dr. Youqi Ke, Prof. Ke Xia, Dr. Eric Zhu, Dr. Dongping Liu, Prof. Xiu Feng Han Youqi Ke, Ke Xia and Hong Guo, PRL 105, 236801 (2010) D.P. Liu, X.F. Han and Hong Guo, PRB 85, 245436 (2012). NEGF-DFT-CPA-NVC 2012-08-29 NEGF5, Jyvaskyla, Finland page 50 MTJ - magnetic tunneling junctions Picture from W. Butler, Nature Mat., 3, 845 (2004) Tunnel barrier is a few atomic layers thick. I tot I tot TMR = I tot Spin transfer torque (STT) Problem: for a given bias, STT is too small, or junction resistance too large. 2012-08-29 NEGF5, Jyvaskyla, Finland page 51 Solution: decreasing the junction resistance Why resistance is large? Because tunnel barrier is an insulator. How to reduce resistance? Dope the insulator with metal atoms. Why it does not work? Because impurity scattering destroys TMR. Youqi Ke, Ke Xia and Hong Guo, PRL 105, 236801 (2010) 2012-08-29 NEGF5, Jyvaskyla, Finland page 52 New idea: Can we find a dopant that exponentially decreases resistance, but only linearly decreases TMR? We thus predict that Zn doping into MgO barrier will solve our problem! D.P. Liu, X.F. Han and Hong Guo, PRB 85, 245436 (2012). 2012-08-29 NEGF5, Jyvaskyla, Finland page 53 Newest: CPA to compute variance by evaluating <GGGG> Eric Zhu & H.G. (2012). Huge device to device variability. F.L. Yang et al., in VLSI Technol. Tech. Symp. Dig., pp. 208, June 2007. 2012-08-29 NEGF5, Jyvaskyla, Finland page 54 Summary By solving 4 critical problems: disorder averaging, band gap, large size and accuracy, NEGF-DFT method can begin to predict device characteristics parameter-free for realistic nanoFET structures. Other details were included into NEGF-DFT as well: electron-phonon, collinear and non-collinear spin, spin-orbit, photon, high frequency, transient, etc. Endless application possibilities… Integration with industry TCAD tools possible. Further reduction of computation time underway … 2012-08-29 NEGF5, Jyvaskyla, Finland page 55 Thank you ! Acknowledgements to Canadian funding: NSERC, CIFAR, FQRNT, IRAP, McGill University. We are grateful to Hong Kong government which funded AoE at HKU where the Sentaurus benchmark was done. 2012-08-29 NEGF5, Jyvaskyla, Finland page 56