JOURNAL OF ADVANCED INSTRUMENTATION AND MEASUREMENT, 2010 1 Self-Tuning Time–Delay Compensation and Equalization for Audio Entertainment Devices T.E. Gibson Y. Matton J.W. Morin D.E. Klenk V.A. Hong Abstract—There are many applications where audio tuning can enhance a users listening experience. The act of tuning an audio setup can be challenging, especially when multiple speakers are involved. To minimize placement difficulties when installing or modifying a speaker setup, an automated system has been conceived. The system automatically adjusts speaker time delay, volume, and frequency content to achieve optimal listening conditions for the specific location of the user. The system sends a binary stochastic signal from each speaker which is then received by two microphones in a remote control. The remote then transmits the recorded microphone data to the receiver to calculate the sound lag and gain for each speaker from the user’s position. The delay and volume of each speaker is then adjusted so that the optimal listening experience may be achieved. Index Terms—time–delay, sound level, equalization, cross correlation, impulse response, system identification I. Introduction T HE position of speakers relative to a listener’s location can dramatically alter sound experiences. Surround sound systems, for example, are designed to stimulate the 3D sphere of human hearing with audio channels above and below the listener. One issue commonly seen with speaker systems is that the intended sound experience can be significantly degraded depending on the location of the listener with respect to the speakers. The reason for this is because without reproducing suggested speaker placement, the time for sound to travel to the user’s ears varies, altering the intended audio profile of the speakers. Additionally, the shape, materials, and existing objects in a room can drastically a↵ect the room’s acoustics. As the audio signal travels from the speaker to the ear, reflection and dampening due to a room’s acoustics may introduce unwanted distortion. With this system, the intent is to minimize the degradation of sound quality through the use of electronics employing signal processing and system identification. The proposed design automatically adjusts the acoustic properties of a speaker system to achieve optimal listening conditions. The design incorporates MEMS microphones into the system remote control to achieve a portable and convenient tuning device. By integrating the microphones into the remote, the system can adjust the system sound properties to be tailored to the listener’s position in the room. While there have been other sound calibration devices in the past, none were designed with the microphones incorporated into the system remote. Another advantage of this design is that it can be used with virtually any type of speaker setup. Existing speaker calibration systems such as the Audyssey R MultiEQTM requires microphones in each speaker, which increases cost and cannot adjust for All authors are with the Mechanical Engineering Department, Massachusetts Institute of Technology, Cambridge MA., 02139, USA a variable listener location. This paper focuses on the system setup and the acoustic characterization methods that are to be performed by the device. Presently, the system prototype utilizes a single speaker and single detection microphone. II. Mathematical Preliminaries: System Identification System identification is a robust field in engineering, devised by control engineers and mathematicians looking to answer the following question. Given the input to a Linear Time Invariant (LTI) system, x(t) where t is time, and knowing the output of that system, y(t), can the dynamics of said system be characterized? This question is illustrated in Figure 1, where h(⌧ ) is defined as the impulse response function which does completely characterize the dynamics of the LTI system. The relationship between the input, the impulse response, and output of a linear system is described below Z 1 y(t) = (h ⇤ x)(t) = h(⌧ )x(t ⌧ )d⌧, (1) 1 where the operation (·) ⇤ (·) is convolution, and in this formulation, h is convolved with x in order to obtain y. In this work, the input and output of the system will be digitally analyzed and for that reason the discrete form of x(t) and y(t) are represented as xi and yi respectively, where the subscript notation indexes the i–th sampled component of the signals. The processes by which the impulse response function is extracted from the known input and output of a system is not straightforward and relies on the concepts of autocorrelation and cross–correlation of signals. The autocorrelation of a discrete time sequence xi of N samples indexed with i = 0 to N 1 is defined as, Cxx,j = N j 1 X (xi N i=0 x̄) · (xi+k x̄) (2) where x̄ denotes the mean of x, j = 0, 1, . . . m and Cxx 2 Rm+1 is a column vector indexed by j with m representing discrete Linear System discrete Fig. 1. Impulse response function in relation to input and output of linear system with discrete approximation for input and output. 2 JOURNAL OF ADVANCED INSTRUMENTATION AND MEASUREMENT, 2010 the total number of lags. The lag time of a system is the inverse of the sampling rate in seconds. As an example, if data was sampled at 1 kHz and m = 1000, then the autocorrelation would contain 1001 data points starting at 0 seconds and spaced in 0.001 second intervals. The crosscorrelation of two finite sequences is defined as Cxy,j = N j 1 X (xi N i=0 x̄) · (yi+k ȳ) (3) where j = 0, 1, . . . m and Cxy 2 Rm+1 is a column vector indexed by j. The cross-correlation and auto-correlation are then related to the discrete impulse response function, h through the following representation of discrete convolution Cxy = Toeplitz {Cxx } h (4) where hj is the j-th lag of the discrete impulse response function h 2 Rm+1 and the Toeplitz operation constructs a matrix of the following form: 2 3 a0 a1 a2 . . . . . . am .. 7 6 .. 6 a1 a0 a1 . . 7 6 7 6 .. 7 . . . .. .. .. 6 a2 a1 . 7 7 (5) Toeplitz{a} = 6 6 . 7 .. .. .. 6 .. 7 . . . a a 1 27 6 6 . 7 . .. a 4 .. a0 a1 5 1 a m . . . . . . a2 a 1 a 0 with a 2 Rm+1 a dummy variable used to illustrate the operation.[1] The discrete impulse response function is then obtained as h = Toeplitz {Cxx } 1 Cxy . (6) Note that inverting the Toeplitz matrix is just the discrete operation for deconvolvinging Cxx from Cxy . The process by which the discrete impulse response function is obtained from discrete inputs and outputs of a linear system have been fully described, however, the computations just described above are computationally burdensome. For that reason the duel of convolution in the frequency domain is introduced, but first the Discrete Fourier Transform (DFT) is formalized. The DFT operation F : CN ! CN is defined as X = F {x} where Xk = N X1 xn e 2⇡i N kn where (·)⇤ is the complex conjugate operator. Recalling that the auto correlation and cross–correlation are convolution operations in the discrete time domain, the following operation is an equivalent definition to the previously defined correlation operations: ⇣ ⌘ Cxx = Real F 1 F{x x̄} · F {x x̄}⇤ (10) ⇣ ⌘ Cxy = Real F 1 F{x x̄} · F {y ȳ}⇤ (11) where the function Real(·) returns the real component of a finite dimensional complex vector. The above defined operation di↵ers from the previous definitions in the scaling of the correlation vectors as well as in their length. Before, in the discrete time domain representation, the total number of lags in the correlations was m + 1, using the Fourier Transform to obtain these correlations the lengths of Cxx and Cxy are N which is equal to the dimension of x and y. In this work the specific functions used to autocorrelate, cross–correlate and deconvolve are contained in Appendix A. Earlier it was mentioned that the discrete time domain operations were costly to compute. Using the DFT as described above will not decrease the computation time required to obtain the impulse response function. However, the Fast Fourier Transform (FFT) is equivalent to the DFT and reduces the number of mathematical operations from O(N 2 ) to O(N log N ). As an example, the FFT of a data set with 10,000 points is 0.4 % of the computation cost when using the DFT. Notice in Appendix A that the FFT function is used and not the DFT function. Fore more details on the FFT refer to [2] and [3] III. Hardware and Experimental Setup A block diagram of the experimental setup is shown in Figure 2. First, a computer is used to generate a stochastic binary signal that is written to an analog output channel of DAQ k = 0, . . . , N 1 Analog Out (7) Analog In n=0 and i is the imaginary unit vector. The inverse Discrete Fourier Transform (iDFT) F 1 : CN ! CN is defined as x = F 1 {X} xn = N 1 2⇡i 1 X Xk e N kn N n = 0, . . . , N 1. Audio Amp Power Supply Mic. Amp (8) k=0 Letting b and c represent finite dimensional complex column vectors, the duel of convolution in the frequency domain: F{b ⇤ c} = F{b} · F {c}⇤ (9) Speaker Fig. 2. Experimental setup. Mic. Gibson, Matton, Morin, Klenk and Hong: TUNING COMPENSATION FOR TIME–DELAY AND EQUALIZATION 3 TABLE I Experimental Equipment. Item Model Manufacturer MATLAB R2009A MathWorks Labview V9.0f3 Nat. Instruments DAQ USB-6215 Nat. Instruments Power Supply E3631A Hewlett-Packard Speaker (15W) TR600-CXi JL Audio Audio Amp LM4755 National Semi. Electret Mic. MD9745APZ-F Knowles Acoustics Electret Amp BOB-08669 Sparkfun.com MEMS Mic.⇤ SPM0404HE5H Knowles Acoustics Op-Amp⇤ LM741CN National Semi. Audio Amp⇤ LM386 National Semi. Speaker (0.5W)⇤ COM-09151 Sparkfun.com *not used in final system a data acquisition unit (DAQ). The analog output signal is fed an audio power amplifier that drives a speaker. Simultaneously, analog input channels on the DAQ sample both the analog output excitation signal and the response signal produced by the microphone. Note, the direct connection between the analog in and analog out is for investigating a delay inherent in the DAQ itself, described below. A power supply provides the power input to the speaker and microphone amplifiers. After the stochastic signal is completed, the computer reads in the recorded sound data from the DAQ’s bu↵er for processing. The experimental apparatus underwent two iterations. The final setup was designed for long range experiments, where the distance between the speakers and receiving microphone approached ten meters. The initial hardware ultimately served as a proof-of-concept as it did not meet range and sensitivity requirements. The initial setup will be outlined below to show briefly the design considerations before focusing on the final setup. Hardware from both the initial and final apparatus designs is listed Table I The initial test setup used a Knowles Acoustics MEMS microphone shown in Figure 3 and a low cost 8 ⌦, 0.5 W speaker (not shown). The MEMS microphone output was amplified by a simple non-inverting amplifier using an LM741 op-amp. The 0.5 W speaker was driven by a low power LM386 audio amplifier. A LabVIEW Virtual Instrument VI was created to interface with the DAQ and MEMS microphone Elecret microphone High-power stereo amp Low-power audio amp Fig. 4. Breadboard with MEMS microphone, Elecret microphone, low–power audio amp, high–power stereo amp and IR sensor. send an excitation signal to the speaker and read in the microphone’s response signal. This setup provided a functional test bed for designing and refining the VI, but it lacked sufficient power for long range tests nor the microphone sensitivity, regardless of the low power output of the speakers. The second and final hardware setup used an electret microphone with its associated signal conditioning circuit, seen in Figure 12. Additionally, a new 15 W speaker and audio amplifier were chosen. The speakers, designed for car stereos, is 4 ⌦ and have a significantly improved frequency response, 59 Hz - 22 kHz, as compared to the previous speaker. To drive the speakers, an LM4755 stereo audio amplifier was used. A photo of the circuits built are shown in Figure 4 IV. Experimental Results Three di↵erent experiments were carried out in this work. Originally two experiments were planned. However, there was an artifact in the data that precipitated a third. The three experiments are presented as follows: DAQ delay investigation, speed of sound verifictaion and frequency shaping. Fig. 3. Microphones Left: Knowles Acoustics MEMS surface mount microphone SPM0404HE5H-PB Right: Knowles Acoustics electret condenser microphone MD9745APZ-F on breakout board, sparkfun.com. A. DAQ Delay While performing experiments to determine the speed of sound in air, a spike in the impulse response function 4 JOURNAL OF ADVANCED INSTRUMENTATION AND MEASUREMENT, 2010 Analog In 1 Analog Out 1 Analog In 2 2 1 h 0 −1 Fig. 5. Analog I/O connections for characterization of DAQ delay −2 appeared at 1 lag. This was repeatable and suggested that there was a delay in the system of 1 lag. In order to confirm this, the analog output from the DAQ was connected directly to the first two analog input channels as shown in Figure 5. This would short circuit the dynamics of the speaker and the microphone and allow the impulse response function of just the DAQ’s D/A and A/D conversion to be extracted. A two second binary stochastic signal was sent from the analog out of the DAQ while sampling on all three channels was performed at 100 kHz. For the first DAQ delay system identification Analog Out 1 was chosen as the input x and Analog In 1 as the output y. The results of this experiment are contained in Figure 6. The autocorrelation function has a value of 1 at 0 lag and negligible magnitude for all other lag values. The autocorrelation function illustrates the memory property of a signal. A signal with an autocorrelation value of 0 for all non–zero lags is a signal whose values have no dependence on previous values. Therefore it is confirmed that the input signal is a random data stream. The cross–correlation function shows a spike at a lag of 1, which is then captured as a spike at a lag of 1 in the impulse response function as well. This confirms the 1 lag delay in the internal dynamics of the DAQ. This experiment was repeated at a sampling rate of 1 kHz with identical results, thus the delay is independent of sampling rate. The experiment was also repeated using Analog Input 2 of the DAQ as y and the 1 lag delay was still present. Cx y Correlation Cx x 1 1 0.5 0.5 0 0 0 5 Lag [1 · 10 10 5 0 s] 5 Lag [1 · 10 10 5 −3 0 50 100 150 Lag [1 · 10 200 5 250 300 s] Fig. 7. Impulse response function for distance of 0.5 m . TABLE II Experimental Data. Distance [m] 4.835 5.317 5.504 Time [s] 0.0136 0.0150 0.0153 One explanation of the delay could be in the choice of the triggering method used in the construction of the LabVIEW Virtual Instruments (VI). The triggering methods for timing of the data acquisition were not all judiciously explored. The true origin of this artifact was never determined. B. Speed of Sound The time delay of the system will be estimated as the lag time of the largest peak in the impulse response function from the audio amp input to microphone amp output as measured by the DAQ. The experimental set up is as depicted in Figure 2 with sampling occurring at 100 kHz. An experimental result of the impulse response function for a speaker and microphone that were 0.5 meters apart is shown in Figure 7. Notice the artifact spike at 1 lag as discussed in the previous section. That spike is ignored and the true delay of the system is estimated as 157 lags (0.00157 seconds). This procedure was repeated at 3 di↵erent distances centered around 5 meters, see Table II. A linear least squares fit was performed on this data to determine the speed of sound. The curve fitting is shown on Figure 8. A value of 346.658 m/s was found for the speed of sound, with an s] 6.5 Distance [m] 0 h −5 −10 6 d = 346.658t + 0.1205 5.5 5 4.5 −15 0 2 4 6 Lag [1 · 10 8 5 s] Fig. 6. Data analysis to determine DAQ delay . 10 12 4 0.013 0.0135 0.014 0.0145 Time [s] 0.015 Fig. 8. Curve fit to determine the speed of sound . 0.0155 0.016 Gibson, Matton, Morin, Klenk and Hong: TUNING COMPENSATION FOR TIME–DELAY AND EQUALIZATION where = 1.4 is the adiabatic index of air, R = 8.314510 J·mol 1 ·K 1 is the molar gas constant, M = 0.0289645 kg·mol 1 is the mean molar mass of air and T is the temperature. Assuming room temperature (293 K) the speed of sound is 346.583 m/s. While the exact temperature of the room was not recorded during the experiment, the low standard deviation, and approximation of the speed of sound to within 1%, assuming T = 293 K, suggest thats this method is accurate at estimating the time it takes for sound to travel in air. C. Equalization: Frequency Shaping (13) with the discrete representation Yp = H · Hp · X. (14) where Yp , Hp , H and X 2 CN . For this experiment a desired output, Ydesired = G !23.5 s2 + 2⇣!1 + !12 !12 (s2 + 2⇣!2 + !22 )1.75 (15) was chosen where G = Y1 , ⇣ = 0.707, !1 = 150 and !2 = 2000 are free design parameters pre–defined to suite the end users desired listening experience, and s = f i with f the frequency and i the imaginary unit vector. The pre– filter was then selected as: Hp , Ydesired /Y (16) where the / symbol denotes point–wise division of two finite dimensional vectors. The frequency shaping experiment was carried out in two steps. First, an unfiltered binary stochastic signal was sent to the audio amp with the microphone 0.3 meters from Fig. 9. Pre–filter for desired frequency content. 10 10 10 10 5 Unfilterred Shaped Log−mean Desired 4 3 2 1 1 10 2 3 10 10 Frequency [1/s] 4 10 Fig. 10. Transfer function and model fit Consider the dynamical block diagram shown in Figure 9. This schematic is similar to Figure 2, however a pre–filter Hp has been placed in front of the plant dynamics H. Note the upper case notation, simply illustrating equivalent dynamics just in the frequency domain instead of time. The goal of the experiment in this section is to design the filter so that the shaped output Yp will have a pre–determined shape in the frequency domain. The above described dynamics are represented as: Yp (f ) = H(f )Hp (f )X(f ), 10 |H(f)| o↵set of 0.1205 m, and a standard deviation of for the linear fit of 0.0001435. Using the ideal gas law, the speed of sound in air can be approximated as r RT c= , (12) M 5 the speaker. The results of this experiment are shown in Figure 10. The light gray lines show the frequency content from the microphone amp, Y . The next step was to determine the pre filter as illustrated in Equation 16. Before the pre–filter could be computed, the noise had to be removed from Y . The black line shows the results from the log–spaced averaging of Y , denoted Ymean . The pre–filter dynamics were then redefined as: Hp , Ydesired /Ymean . (17) The second component of the experiment was then carried out by passing the binary stochastic signal through the pre–filter before entering the audio amp. The results from this experiment are also contained in Figure 10. The dash–dot line represents the desired frequency content of the microphone output and the dark gray line the actual filtered output. From the figure it is apparent that the unfiltered and filtered microphone outputs are distinctly di↵erent. The filtered system response is similar to the desired frequency shape. V. Hand-Held Interface Design The experiments just described were a proof of concept for a handheald device. The final product would center around a remote control with two microphones imbedded in the side, Figure 11. A low–power RF transceiver (e.g. TI-CC2500) would serve as a data link between the microphone and audio receiver. An ARM R microcontroller interfaced with a transceiver will be integrated inside the remote, and a similar microcontroller-transceiver architecture would preside in the audio receiver. In this stated configuration: (1) the audio receiver would transmit a binary stochastic signal to each speaker (2) the acoustic profile of each speaker would be received by the microphones (3) the recorded information would then be transmitted back to receiver were the time delay would be determined as well as the frequency signature of each speaker (4) each 6 JOURNAL OF ADVANCED INSTRUMENTATION AND MEASUREMENT, 2010 Appendices A. Matlab Scripts Used In the following MATLAB scripts, the frequency techniques for convolution were utilized following [4]. A. Autocorrelation function Fig. 11. Concept remote with two electret microphones . speaker would then have a unique time delay and pre–filter setting within the receiver, ensuring an optimal listening experience. VI. Conclusions and Future Work The audio calibration system can achieve optimized sound conditions with features not currently availed in existing products. To design this system, software packages such as LabVIEW R and MATLAB R have provided quick and efficient environments for testing signal processing algorithms, which can be ported over to specific embedded hardware in future iterations. While the system currently employs a single microphone and speaker, it can easily be adapted for listening environments containing multiple speakers. With a user-friendly interface connected to high-quality audio components, the SpeakerBox audio adjustment system will have the capability of delivering an excellent listening experience through the use of delay adjustment and frequency adjustment. VII. Acknowledgements This work was completed in fulfillment of the final project in course 2.131 Advanced Instrumentation and Measurement at the Massachusetts Institute of Technology. The authors would like to thank Professor Ian Hunter for illustrating the techniques for proper system identification with specific recognition for his help on developing the frequency shaping techniques used in this work. The authors are also extremely grateful for the help that the teaching assistant, Adam J. Wahab, o↵ered. Adam was instrumental in the formulation of the project idea and helped with the purchase and design of all the hardware components. A special thanks to National Instruments for the DAQ and LabVIEW software that was used. References [1] Ian W. Hunter, “2.131 advanced instrumentation and measurement,” Class Notes, Spring 2010. [2] J. W. Cooley and J. W. Turkey, “An algorithm for the machine computation of the comlplex fourier series,” Mathematics of Computation, vol. 19, 1965. [3] P. Duhamel and M. Vetterli, “Fast fourier transforms: A tutorial review and a state of the art,” Signal Processing, vol. 19, 1990. [4] G.E.P Box, G.M. Jenkins, and G.C. Reinsel, Time Series Analysis: Forecasting and Control, Prentice-Hall, 3rd edition, 1994. ------------------------------------------------------function ACF = autocorr(Series , nLags) nFFT =2^(nextpow2(length(Series)) + 1); F =fft(Series-mean(Series) , nFFT); F =F .* conj(F); ACF =ifft(F); ACF =ACF(1:(nLags + 1)); ACF =ACF ./ ACF(1); ACF =real(ACF); ------------------------------------------------------- B. Crosscorrelation function ------------------------------------------------------function XCF=crosscorr(Series1, Series2, nLags) Series1 =Series1 - mean(Series1); Series2 =Series2 - mean(Series2); L1 =length(Series1); L2 =length(Series2); nFFT =2^(nextpow2(max([L1 L2])) + 1); F =fft([Series1(:) Series2(:)] , nFFT); XCF =ifft(F(:,1) .* conj(F(:,2))); XCF =XCF([(nLags+1:-1:1)(nFFT:-1:(nFFT-nLags+1))]); XCF =real(XCF) / (sqrt(ACF1(1)) *sqrt(ACF2(1))); ------------------------------------------------------- C. Deconvolution function ------------------------------------------------------function [q,r]=deconv(b,a) [mb,nb] = size(b); nb = max(mb,nb); na = length(a); if na > nb q = zeros(superiorfloat(b,a)); r = cast(b,class(q)); else [q,zf] = filter(b, a, [1 zeros(1,nb-na)]); if mb ~= 1 q = q(:); end if nargout > 1 r = zeros(size(b),class(q)); lq = length(q); r(lq+1:end) = a(1)*zf(1:nb-lq); end end ------------------------------------------------------- Gibson, Matton, Morin, Klenk and Hong: TUNING COMPENSATION FOR TIME–DELAY AND EQUALIZATION 0 1 2 3 0 1 2 3 VCC 4 5 6 7 5 7 6 VCC 24V A 4V 4 8 A A 1000µF 10k VCC VCC VCC 6 Microphone 7 D 2.7k 100k 3 E 4 5 2 135k Mute LM4755 2.7 0.1µF A 5 10µF F 10V 8 E 12k A VCC 7 5V Fig. 12. Electret microphone breakout board amp arrangement using OPA344. E 6 1000µF 8 AmpA + 0.1µF SIG INA 2 0.1µF 83k 100pF 1 + Bias 100µF D 0 2.7 C GND 83k - 0.1µF 0.1µF HDR1X3 D SIG C 1 1k 1 AmpB 3 INB 2 4 1000µF 2.7k OPA344 5 3 1µF 135k B 10k 2.2k C 4 B GND 1µF 4V - B VCC 4V 4V 4V F F 6 B 250µF LM386 5 Speaker Je↵rey W. GMorin (BS’09) was born in Manchester, New Hampshire, in 1986. He 0 1 2 4 5 received the 3B.S. degree mechanical C in engineering from the University of New 6 7 8 Hampshire, Durham, in 2009. He is currently a M.S candidate at Massachusetts Institute of Technology, Cambridge, where he is researching active fluids and robotic locomotion under the supervision of Dr.D Anette Hosoi. In the summers of 2005 to 2009 he worked at Advanced Combustion Technology, Hooksett NH, where he developed ultra-low NOx burners for large scale power plants. He is also a student member of Tau Beta Pi and ASME. + 3 4 C 1 G 2 SIG 10k 0 Fig. 14. High power stereo amp arrangement usingB LM4755. 8 0.05µF 7 G 1 2 3 4 - 10 5 GND Vin D Fig. 13. Low power audio amp using LM386. E B. Circuit Diagrams E Dan E. Klenk (S.B.’09) received an S.B in F mechanical engineering from the Massachusetts Institute of Technology, Cambridge, in 2009. He is currently a master’s student also at MIT researching walking dynamics and biomechanics. F Travis E. Gibson (B.S.’06 M.S.’08) was born in Jacksonville, FL, in 1984. He received the B.S. degree in mechanical engineering from Georgia Institute of Technology, Atlanta, in 2006 and the M.S. degree in mechanical engiG neering from Massachusetts Institute of Technology, Cambridge in 2008. He is currently a Ph.D. candidate at MIT, where he is researching adaptive control under the supervision of 0 1 2 3 4 Dr. Anuradha Annaswamy. In the summers of 2003 to 2006 he worked at Vistakon, Jacksonville FL, summers of 2008 and 2009 at NASA Langley Research Center, Hampton VA, and Summer 2010 at Boeing, Huntington Beach CA. He is also a student member of the AIAA and IEEE. Yves Matton (B.S.’08) was born in Paris, France, in 1986. He received the B.Sc. degree in technological innovation from Ecole Polytechnique (Paris) in 2008, and is currently a M.S. degree candidate in mechanical engineering at MIT. He is currently research assistant working on new newtonian fluids under the supervision of Professor L.J Gibson and Professor G.H.McKinley. G 5 6 7 8 Vu A. Hong (S.B.’10) was born in San Diego, California, in 1988. He is currently an S.B degree candidate in mechanical engineering at the Massachusetts Institute of Technology, Cambridge, where he is researching bacteria mediated ice nucleation under the supervision of Professor Evelyn Wang. He will be working at Palo Alto Research Center, Palo Alto, CA, this summer, and will be pursuing graduate studies at Stanford University thereafter. 6