A Photoelectric-Stimulated MoS2 Transistor for Neuromorphic Engineering

The von Neumann bottleneck has spawned the rapid expansion of neuromorphic engineering and brain-like networks. Synapses serve as bridges for information transmission and connection in the biological nervous system. The direct implementation of neural networks may depend on novel materials and devices that mimic natural neuronal and synaptic behavior. By exploiting the interfacial effects between MoS2 and AlOx, we demonstrate that an h-BN-encapsulated MoS2 artificial synapse transistor can mimic the basic synaptic behaviors, including EPSC, PPF, LTP, and LTD. Efficient optoelectronic spikes enable simulation of synaptic gain, frequency, and weight plasticity. The Pavlov classical conditioning experiment was successfully simulated by electrical tuning, showing associated learning behavior. In addition, h-BN encapsulation effectively improves the environmental time stability of our devices. Our h-BN-encapsulated MoS2 artificial synapse provides a new paradigm for hardware implementation of neuromorphic engineering.

Here, we demonstrate an efficient photoelectrical tunable h-BN-encapsulated MoS 2 synaptic transistor with basic synaptic functions. Furthermore, under electrical modulation, we successfully simulate the impressive Pavlov classical conditioning experiment through V bg tuning, which realizes the acquisition, extinction, and recovery function of associated learning. Due to the h-BN encapsulation, our devices exhibit superior environmental time stability. Our h-BN-encapsulated MoS 2 artificial synaptic transistor provides a novel paradigm for neuromorphic engineering based on 2D materials.

Results
First, we fabricated an h-BN-encapsulated MoS 2 synaptic transistor on an AlO x /Si substrate, which simulates synaptic behavior by photoelectric stimulation, as shown in Figures 1(a) and 1(b). 2D layered h-BN and MoS 2 were prepared by mechanical exfoliating. The surface morphology of our device was characterized by scanning electron microscopy (SEM) and atomic force microscopy (AFM), as shown in Figure 1(c) and Figure S2a, respectively, showing a typical channel width of 10 μm, a length of 15 μm, and the thickness of the MoS 2 ; h-BN was approximately 1.7 and 7 nm. The Raman spectrum shows the characteristic peaks of both materials: Raman shift of the MoS 2 characteristic peak is 385,405 cm -1 (Figure 1(e)) and the h-BN is 1366 cm -1 ( Figure S2b), which is consistent with previous reports. Figure 1(d) shows the Raman mapping of h-BN-encapsulated MoS 2 synaptic transistor at 405 cm -1 , the channel MoS 2 exhibits intense intensity, and the h-BN/MoS 2 overlap region is more strongly correlated with h-BN encapsulation, where the black and gray dashed areas represent the h-BN/MoS 2 overlap region and channel MoS 2 , respectively. A significant peak was observed in the PL spectrum of MoS 2 at 1.88 eV photon energy (Figure 1(f)), which is consistent with the band gap of multilayer MoS 2 . Then, we studied the behavioral characteristics of our h-BN-encapsulated MoS 2 synaptic transistor under electrical modulation. Figure 2(a) shows the I ds -V bg curves of the h-BN-encapsulated MoS 2 synaptic transistor with V ds of 0.1, 0.5, and 1 V. The back gate voltage was swept from -6 to 8 V, then swept back, and a noticeable clockwise hysteresis loop was observed, which may be due to charge trapping between the MoS 2 and AlO x interfaces. The statistical distribution of the maximum value of the memory window indicates that the memory window of most devices is 2~3 V (see the statistics of 80 devices in Figure S11a in Supplementary Materials). The transfer curves of the h-BN/MoS 2 /h-BN control devices show no hysteresis window, since the bottom h-BN isolates the interface effect of MoS 2 and AlO x (see Figures S10a-c, the schematic diagram of the control devices, micrograph of the control device, and transfer curves of the control devices in Supplementary Materials). Owing to top encapsulated h-BN, the stability of our devices has been significantly improved (see Figure S3, output curves and stability of h-BN-encapsulated MoS 2 synaptic transistor in Supplementary Materials) [52][53][54][55]. We explored the optimal base and pulse voltages for device operation in electrical mode for excitatory and inhibitory synapses, with reference to gain (A 5 /A 1 , the amplitude of the postsynaptic current caused by spike is denoted by A) of five consecutive pulses and long-term synaptic weight changes (ΔW/W, calculated by ðI − I0Þ/I0 * 100%, where I0 and I represent the current states before and after the application of the pulse signal, respectively. Before applying the pulse signal, we select the average value at the 5th second as I0. After the pulse signal is applied, the average value of the 40th second is selected as I, and the pulse signals are applied at the same time. For excitatory synapses, the gain was maximized when V bg base was -3 V and pulse was -4 V (pulse duration of 10 ms, interval of 200 ms), as shown in Figure 2(b) (no synaptic excitability of the h-BN/MoS 2 /h-BN control devices under the same V bg base and pulse conditions, see Figure S10d in Supplementary Materials). For inhibitory synapses, excitatory spike stimulation was first performed, and then fixed base, incremental V bg pulse was applied, and gain and weight changes were reduced, that is, the depression effect gradually strengthened, and 8 V was selected as the inhibitory spike (duration of 10 ms, interval of 200 ms), as shown in Figure 2 Figure S4 in Supplementary Materials. The electrical potentiation and inhibition effects under electrical stimulation are attributed to the charges trapping and detrapping at the MoS 2 -AlO x interfaces. The statistical distribution of the maximum value of the excitatory index indicates that the excitatory index of most devices can reach 500-700% (see the statistics of 80 devices in Figure S11b in Supplementary Materials). Under forward bias (V bg pulse of 8 V), the oxygen vacancy trapping states in AlO x move toward the channel, trapping the electrons in MoS 2 , causing channel current to decrease, corresponding to synaptic inhibition. While under reverse bias (V bg pulse of -4 V), oxygen ions in AlO x move toward MoS 2 , and the oxygen vacancy trapping states release trapped electrons, resulting in increased channel current, which corresponds to synaptic potentiation (see Figure S5, physical mechanism under electrical stimulation in Supplementary Materials).
The realization of the association learning is of great significance for neuromorphic engineering. Pavlov's dog classical conditioning experiment is a typical associative learning experiment in physiology [56][57][58]. In Pavlov's dog experiment, food is called unconditional stimulation (US), while the bell and salivation are called neutral stimulation (NS) and unconditional response (UR), respectively. Food can cause salivation, while bell ringing alone does not cause salivation. Combining the bell with food, that is, after the bell rings, the dog is fed with food, also causes salivation [57,59]. Pavlov's dog classical conditioning experiment can be simulated on the proposed h-BN-encapsulated MoS 2 synaptic transistor by efficient electrical modulation, as shown in Figure 3. V bg (base, pulse) of (-5, -4 V) applied to the presynaptic gate is considered to be "bell" (NS), and V bg (base, pulse) of (-3, -4 V) is considered "food" (US). The postsynap-tic source drain channel current acts as synaptic weight, and the synaptic weight of 20 nA is defined as the threshold for the "salivation" response (UR). After a single training, only the "bell" ringing does not cause salivation, but after repeated training, the "bell" ringing can also cause "salivation," which shows the same effect as feeding "food." At this point, an association is established between "bell" and "food," and    V bg (base, pulse) of (-5, -4 V) applied to the presynaptic gate is considered to be "bell" (NS), and V bg (base, pulse) of (-3, -4 V) is considered "food" (US). The postsynaptic source drain channel current acts as synaptic weight, and the synaptic weight of 20 nA is defined as the threshold for the "salivation" response. the corresponding NS "bell" is converted to conditional stimulation (CS), causing a conditional response (CR) that triggers "salivation" similar to US, which is called acquisition. After a long time or reset operation, "salivation" no longer occurs when there is only CS, which means that the association between CS and US is extinct/forgotten. However, after training again, "salivation" occurs again when the "bell" rings only, that is, the association is recovered. In addition, we found that due to the existence of acquisition, the current of single training after recovery is significantly higher than the previous single training, which has exceeded the threshold and "salivation" occurs.
In addition to electrical modulation, optical spikes also enable efficient regulation of our h-BN-encapsulated MoS 2 synaptic transistor, which uses laser pulses as the photogate to adjust the channel conductance (synaptic weight), as shown in Figure 4(a). Figure 4(b) shows the single-laser pulse characteristics (532 nm, duration of 100 ms, power of 50 mW/cm 2 ) of the synaptic transistor at V bg of 0, -5, and -10 V, which significantly affect the reference current (the single-laser pulse characteristics of our synaptic transistor at 473,655 nm and the single-laser pulse current ver-sus time at three wavelengths with V ds of 1 V are detailed in Figure S6 in Supplementary Materials). Besides, variation of postsynaptic current amplitude under different V bg (0, -5, -10 V) and single-laser pulses with different wavelengths (473, 532, 655 nm) is shown in Figure 4(c). We found that the PSC amplitude increases significantly with V bg , where different wavelengths have little effect on the PSC amplitude (both of μA), which may be due to the excitation of the h-BN-encapsulated MoS 2 synaptic transistor at each wavelength, resulting in photocarrier accumulation in the channel. Specifically, photogenerated carriers (electronhole pairs) are generated and separated in the top h-BN under laser duration, in which photogenerated electrons are transferred to MoS 2 , resulting in an increase in channel current. With the cumulative number of laser pulses, the electrons in MoS 2 increase continuously, and the channel current appears to be nonvolatile, corresponding to the LTP behavior of neural synapses (see the physical mechanism under optical stimulation in Figure S9 in Supplementary Materials). Moreover, paired pulse facilitation (PPF) is a dynamic increase in neurotransmitter release that is thought to be critical in biosynaptic function simulations [60], where presynaptic-induced EPSC amplitude decreases with increasing two consecutive pulse intervals (Δt).  [16,20,[62][63][64][65]. Moreover, we demonstrate long-term synaptic potentiation and inhibition effects under photoelectric modulation with V bg and V ds of 0 and 1 V, i.e., implementing optical potentiation (532 nm, laser duration 100 ms, power 50 mW/cm 2 ) and electrical inhibition (V bg pulse 3 V, duration 50 ms) behaviors in sequence, as shown in Figure 4(f). The optimal V bg pulses for inhibition under optical stimulation are explored in Figure S7 in Supplementary Materials, and 50 laser-stimulated LTP, followed by 50 electrical stimulation LTD characteristics, are shown in Figure S8 of Supplementary Materials, respectively. PPF Subsequently, we demonstrate the optical neural plasticity of the h-BN-encapsulated MoS 2 synaptic transistor. Figure 5   are 100 ms and 1 s, power of 50 mW/cm 2 , laser number of 50, V bg and V ds are 0 and 1 V), and Figure 5(b) is an enlargement of the dotted circle region in Figure 5(a). Gain (A n /A 1 ) variation of different wavelengths (473, 532, 655 nm) and pulse numbers under laser stimulation (532 nm, laser duration and interval are 100 and 400 ms, power of 50 mW/cm 2 , V bg and V ds are 0 and 1 V) are demonstrated in Figure 5(c), which accumulates the laser pulse numbers, and the gains under three wavelength stimuli gradually increase and tend to saturate. Besides, we demonstrate that with the increase of the laser spiking number, the long-term synaptic weight changes at different wavelengths also gradually increase, indicating the synaptic connections are strengthened, as shown in Figure 5(d).
And we found that the ΔW/W induced by the 532 nm laser spike is the largest, that is, the strongest synaptic connection strength, and the weakest was at 655 nm, which may be attributed to the larger the wavelength, the smaller the energy under the same conditions, and the fewer photogenerated carriers are generated, resulting in the weakest connection strength. However, 532 nm may more easily excite our h-BN-encapsulated MoS 2 synaptic transistor than 473 nm, by concentrating more photogenerated carriers. Figure 5(e) shows the gain (A 50 /A 1 ) variation of different wavelengths and laser powers under optical modulation (laser duration and interval are 100 and 400 ms, V bg and V ds are 0 and 1 V). Abnormally, as the laser power increases, the synaptic gain decreases, which may be attributed to the incremental power intensity causing slight damage to the channel material and degradation of performance. Finally, we also demonstrate the synaptic gain as a function of wavelength and laser frequency (1-50 Hz, duration of 100 ms, power of 50 mW/cm 2 , V bg and V ds are 0 and 1 V), which increases with frequency and has a maximum at 50 Hz, as shown in Figure 5 Table S1 in Supplementary Materials). The acceptable switching power consumption is estimated to be 80 pJ per spike, which is two orders of magnitude lower than the traditional CMOS [66] and even down to femtojoule when the V ds is 0.1 V, close to the human brain [16]. The h-BN-encapsulated MoS 2 artificial synaptic transistor provides a novel paradigm for neuromorphic engineering based on 2D materials.

Discussion
In conclusion, our breakthrough, efficient, photoelectrical tunable, diverse h-BN-encapsulated MoS 2 synaptic transistor demonstrates basic synaptic functions including EPSC, PPF, LTP, LTD, synaptic gain, frequency, and weight plasticity. In addition, under electrical modulation, we successfully simulated the Pavlov classical conditioning experiment and realized the associated learning function. It is worth mentioning that due to the h-BN encapsulation, our devices have superior environmental stability.
Our synaptic transistor provides an unparalleled perspective on novel 2D material-based neuromorphic engineering and brain-like computing.

Preparation of the h-BN-Encapsulated MoS 2 Synaptic
Transistor. We fabricated the h-BN-encapsulated MoS 2 synaptic transistor on an AlO x /Si substrate, which simulates synaptic behavior by photoelectric stimulation, as shown in Figures 1(a) and 1(b). Firstly, two-dimensional layered h-BN and MoS 2 were prepared by mechanical exfoliation, and their thicknesses were determined by an atomic force microscope (AFM) to be about 7 and 1.7 nm, as shown in Figure S2a.  Figure 1(

Conflicts of Interest
The authors declare no conflict of interest. With the cumulative number of laser pulses, the electrons in MoS 2 increase continuously, and the channel current appears to be nonvolatile, that is, LTP behavior [11]. Figure S10:  Table S1: comparison of 2D-based synaptic device performance, including device geometry, operating modes, electrical/optical tuning, excitatory index, long-term weight change, and power consumption. (Supplementary Materials)