Supplementary Material for Synaptic Computation Enabled by Joule Heating of Single-layered Semiconductors for Sound Localization Authors: Linfeng Sun 1, Yishu Zhang 2, Geunwoo Hwang 1, Jinbao Jiang 1,3, Dohyun Kim 1, Yonas Assefa Eshete 1, Rong Zhao 2 *, Heejun Yang 1 * Affiliations: 1 Department of Energy Science, Sungkyunkwan University, Suwon 16419, Korea 2 Singapore University of Technology & Design, 8 Somapah Road, 487372, Singapore 3 IBS Center for Integrated Nanostructure Physics (CINAP), Institute for Basic Science, Sungkyunkwan University, Suwon 16419, Korea These authors contributed equally to this work. *Correspondence to: zhao_rong@sutd.edu.sg, h.yang@skku.edu This PDF includes: Materials and Methods Figs. S1 to S14.
Materials and Methods Details for CVD Growth of Monolayer MoS2: 1) The monolayer MoS2 covered by bilayer islands in this work (Group 1 mentioned in Fig. S4) was grown on a clean silicon wafer covered with a 300 nm SiO2 layer. Molybdenum trioxide (MoO3) and sulfur (S) powder were used as solid precursors. The experimental setup is shown below. Ceramic boats with MoO3 and S powder were located at different temperature zones. The growth temperature was 800 C, which was increased by a rate of 20 C min -1 from room temperature, kept for 30 min. Then, the furnace was cooled down naturally. Before increasing the temperature for growth, 200 sccm argon gas was flowed inside the furnace for approximately 15 min to achieve a pure argon atmosphere, and the gas flow was decreased to 50 sccm during the growth process. Gas Vacuum chamber SiO 2 wafer Exhaust S powder MoO 3 powder Experimental setup for the growth of monolayer MoS2 with bilayer islands (Group 1). 2) Another group of monolayer MoS2 (Group 2 mentioned in Fig. S14) was synthesized using a mixture of ammonium heptamolybdate (AHM, Sigma-Aldrich, 431346), NaOH and an iodixanol solution (Sigma- Aldrich, Opti Prep, D1556) as precursors. AHM (0.2 g) was dissolved into 30 ml of deionized (DI) water (solution 1), and NaOH (0.1 g) was dissolved into 30 ml of DI water (solution 2). Then, solution 1 and solution 2 were mixed together with the Opti Prep solution in a volume ratio of 0.3:3:0.5, followed by a two-step spin coating (500 rpm for 5 s first and then 3000 rpm for 45 s) processes on the prepared clean SiO2/Si wafer (SiO2: 300 nm). Then, the wafer was sent into a two-zone furnace for sulfurization (zone 1), and the sulfur powder (0.75 g) is placed in zone 2. First, the precursor was annealed at 500 C for 10 min under vacuum, and then, the temperature was increased to 800 C for zone 2 and 200 C was kept for zone 1 for 20 min under an argon gas atmosphere (gas flow: 500 sccm). Then, H2 (5 sccm) was introduced to act as a reducing agent for 3min. Finally, the furnace was cooled down naturally. Vacuum chamber Gas Zoom 1 Zoom 2 Exhaust S powder SiO 2 wafer with precursor Experimental set up for the growth of monolayer MoS2 (Group 2).
Optical Characterization: Raman and PL spectra were measured on a Witec Alpha300 Confocal System located in a glove box, which was filled with inert gas (argon), and the concentrations of O2 and H2O were 0.6 ppm and 0 ppm, respectively. The excitation wavelength for both the Raman and PL measurements was 532 nm, and the gratings used for the Raman and PL measurements were 1800 and 600 mm -1, respectively. The integration times for the Raman and PL images were 0.2 s and 0.1 s per data point, respectively. The integration times for the Raman and PL spectra were 20 s, and 10 s, respectively. The laser power was controlled below 0.5 mw to avoid the heating of the samples during the measurements. A 100x objective (Zeiss) was used to focus the laser beam. Device Fabrication: The substrate with monolayer MoS2 (Group 1) was first spin-coated with poly(methylmethacrylate) (PMMA) 950 A4. The initial rotation speed was 500 rpm for 5 s, and, the speed was increased to 4000 rpm for 60 s. Then, the sample was baked at 180 C on a hot plate for 2 min. Next, electron beam lithography (EBL) was used to expose the electrode pattern. Finally, the sample was metallized with Cr/Au (5/50 nm), followed by a 2 h lift-off process in acetone. For the Group 2 sample, the device fabrication process was the same as that for the Group 1 sample, except it was transferred onto the same SiO2/Si substrate before the device fabrication process. MoS2 nanoribbons were etched by SF6 plasma (power: 30 W; time: 30 s). Sound localization: In the manuscript, we demonstrate a key function for the realization of sound localization: how to use synaptic computation to take only interaural time difference (ITD) information by suppressing interaural level difference (ILD) signal, which is vital for sound localization. We used low frequency signal inputs in this work (Fig. 3c and 3d), because such a binaural effect becomes protruding only when the spike frequency from sound source is less than 1000 HZ (4 th edition, neuroscience, printed in USA), and in this frequency range, the effect of sound localization highly depends on the interference of ILD and ITD. The detailed information for the realization of sound localization is as follows: In human auditory system, the sound is encoded by the auditory receptor and hair cell (cochlea), located at inner ear, and the encoded signal is transferred into a series of electric spike patterns through cochlea nuclei. Once the sound is transferred into the electric signal, we can realize sound localization by using only electric signal. We note that the sound encoding into electric spike is not an issue for sound localization or other neuromorphic technology and the signal processing (that we show in this work) has been the bottleneck of the technology. Here, we used the spike patterns as our inputs, mimicking the biological system. The input spike trains can be interpreted as sound signal. The neurons and synaptic plasticity in Fig. 3 are required as biological constraints. Otherwise, we cannot process signals by frequency-selective excitatory (inhibitory) synapses. Furthermore, without synaptic plasticity, the circuit cannot go back to the initial state. Here, we adopt the Jeffress model to explain how auditory systems in our brain use the time difference between the arrival times of sounds spikes from two ears (cochleas) to find the location sound sources (see the figure
below). Both intensity and time differences exist between two spike trains from two ears. In Fig. 3b, we demonstrated a function of eliminating the intensity interference (ILD). This makes the spikes enter the MSO only with the time delay effect (ITD) from two ears. Depending on the time difference, specific neurons make a fire, which tells the brain the location of the sound source; below, we demonstrate that different neurons can be activated by different time differences (ITD) by the Jeffress model. Jeffress model for binaural psychophysical observations. In order to implement Jeffress model, we design a simulated circuit which uses pulse delay circuits as the delay lines, and use and gates as the coincidence detectors. Once the two pulses trains with fixed ITD cues after filtered by the circuits shown in Fig. 3b, the coincidence detector (neuron) which has the exact delay time for such a ITD cue will make a fire, as shown in Figure (with time delay circuits) below. P1, P2, and P3 represent three different azimuthal angles (θ=0, 90, and 180 ) shown in Fig. 3b. The detector for the selective output (neuron fire) when the sound directions comes from the position with different azimuthal angles (θ=0, 90, and 180 ).
The time delay circuit used above are implemented by two mono-stable retriggerable chips. As shown below, the input pulses are delayed by the charging and discharging process in RC circuit. The delay time is controlled by the formula, Td=R1*C1. By adjusting the values of R1, C1, the different delay time can be realized. Time delay circuit used for demonstration the ITD function. Kelvin Probe Force Microscopy Measurement: Kevin probe force microscopy (KPFM) can show the surface potentials of materials and has wide applications in characterizing the surface electronic properties of materials. KPFM is a modified AFM technique, which can simultaneously obtain information on the topography and surface potential. A voltage feedback loop is applied between the tip and the sample to monitor the oscillation caused by the contact potential difference between the tip and sample. In our measurement, HITACHI SI-DF3-R (K-A102002995) rhodium (Rh)-coated Si with a force constant of 1.3 Nm -1 was used. The resonance frequency of the cantilever was 25 khz. A KPFM image of the sample surface was acquired with the cantilever vibrating at a frequency of 25.2 khz, which was almost the same as its resonance frequency. The scan rate of the probe was 0.5 µm/s, and the DC voltage of the tip was set to 5 V. The experimental setup of KPFM
a Conductance (ns) 0.3 Read at 0.1V 0.2 Initial Conductance 0.1 0 2 4 6 8 10 Cycle Index 0.0-0.2-0.4-0.6-2.0-1.5-1.0-0.5 0.0 Voltage (V) Figure S1. STP behavior of the MoS2-based two-terminal device. a. Device conductance after each voltage sweep shown in Figure 1c with a reading voltage of 0.1 V. The first data point represents the initial device conductance. b. Conductance plasticity trend with negative voltage bias. The variation of conductance is smaller than that of the positive voltage sweeps. The arrow represents the direction of the current increased (same as Figure S2, S7 and S14). b Current (na)
Current (na) c a Current (na) 0-10 -20-30 -40-50 -2.0-1.5-1.0-0.5 0.0 Voltage (V) 0.0-0.2-0.4-0.6-2.0-1.5-1.0-0.5 0.0 Voltage (V) b Current (na) d Current (na) 3.2 2.4 1.6 0.8 0.0 0.0 0.5 1.0 1.5 2.0 Voltage (V) 0.03 0.02 0.01 0.00 0.0 0.5 1.0 1.5 2.0 Voltage (V) Figure S2. Temperature-dependent Id-Vd curves under voltage sweeps. Conductance plasticity for the negative (a) and positive (b) voltage sweeps at room temperature (300 K). At 15K, conductance plasticity for negative (c) and positive (d) voltage sweeps.
Figure S3. Four-probe resistance measurement. a, Optical image of the device used for the measurement. b, Comparison of the I-V curves between the two-probe and four-probe measurements.
Time Constant (ms) 10k 1k 100 10-10 -5 0 5 10 Gate Voltage (V) Figure S4. Gate-tunable time constants of STP, in consistent to Figure 1e. The time constant represents the time for the synaptic strength to decay to 1/e of its original value. When the applied gate voltage is 10 V, there is no STP (Figure 1e). The recovery time constant with a gate voltage of 10 V for the synaptic strength was too long to be measured, which is displayed as 10k in the above figure.
Figure S5. Operation energy of synaptic device. a, SEM image of device used for the calculation of operation energy. b, The current of monolayer-mos2-nanoribbon-based device used for measuring the energy consumption with the applied pulse voltage shown in Figure 1f.
Post-synaptic synaptic current A1 A2 Pre-synaptic pulse Δt 1 st pulse 2 nd pulse Figure S6. Stimulus of paired pulses used for characterizing PPF (A1<A2) and PPD (A1>A2). A1 and A2 are the absolute amplitudes of the excitatory post-synaptic currents (EPSCs).
a Current(μA) Figure S7. 5 4 3 2 1 0 0.0 0.5 1.0 1.5 2.0 Voltage (V) b Resistance (MΩ) 4 3 2 1 0 0 2 4 6 8 10 Time (s) Mimicking short-term depression based on heavily n doped material (Vg=20V). a, Resistance (conductivity) plasticity behaviors under bias voltage sweeps. The resistance increases after each voltage sweep. b, Resistance recovery and automatic decay. The read voltage is set as 0.1 V.
Steady State Ratio 8 6 4 2 WSC WOSC 0 10 20 100 150 200 Stimulus Rate (Hz) Figure S8. The steady-state EPSC amplitude ratio, defined as the ratio of the 10th EPSC to the first EPSC, is kept constant in a wide frequency range with our synaptic computation for ITD-based sound localization. WOC and WOSC means that with synaptic computation and without synaptic computation, respectively.
Left input Right input 0º neuron 90º neuron 180º neuron P1 P2 P3 Figure S9. The output signal from MSO neurons in three cases (θ = 0, 90, and 180 ) under the geometry described in Figure 3b. P1, P2, and P3 represent three different angles (0, 90, and 180 ) shown in Figure 3b. The left (right) input represents the signal from left (right) CN after encoding by the circuit shown in Figure 3.
Figure S10. Surface characterization of the monolayer MoS2 used in this work. a, Optical image of a CVD grown monolayer MoS2 sample with bilayer islands. b, AFM image of the zoomed-in area for the smaller bilayer islands shown in (a).
a ML BL b ML BL Intensity (a.u.) Intensity (a.u.) 600 640 680 720 Wavelength (nm) 360 375 390 405 420 Wavenumber(cm -1 ) Figure S11. Optical spectra of the monolayer MoS2 sample. a, PL and b, Raman spectra of two different areas in the sample: monolayer (ML), and bilayer (BL).
Figure S12. Device fabrication and KPFM measurement. The optical image of the device used for KPFM.
Figure S13. Charge transfer between the first and second layer in MoS2. a, Schematic diagram for the charge transfer process between two adjacent layers. The bilayer region attracts electrons from the surrounding monolayer region. b, Band alignments of the monolayer, bilayer, and bilayer islands. Both Fermi levels for the bilayer and bilayer islands are lower than that for monolayer MoS2.
Figure S14. Opto-electrical characterization of pure monolayer MoS2 for comparison. a, Optical image of the device based on pure monolayer MoS2. b and c, PL and Raman mapping images, respectively. The sample is uniform, which can be seen from the Raman and PL mappings. d, Raman and PL spectra of the monolayer sample. e and f, Conductivity plasticity behaviors under negative and positive voltage sweeps, respectively.