Introduction

In the current era of big data, brain-inspired neuromorphic computing featuring parallel data processing is considered promising technology to efficiently process considerable unstructured data, such as texts, images, audio, and video1,2,3,4,5,6,7. Therefore, it is necessary to develop hardware neural networks (HNNs) as a platform to support the energy-efficient operation of neuromorphic computing. This is because the existing von Neumann architecture typically functions as a hardware platform for conventional serial computing and is, therefore, unsuitable for neuromorphic parallel computing8,9,10,11,12,13. Various approaches employing digital and analog circuit technologies have been suggested to implement the HNNs having learning algorithm functionality8,14. Neural networks generally consist of neurons and synapses, where each neuron communicates with other neurons via synapses. Previous studies proposed circuit technologies that artificially emulated the functionalities of synapses and neurons and implemented algorithm-based learning abilities14. Recent studies have focused on implementing HNNs with a combination of advanced memory devices for synapses and peripheral analog circuits for neurons6,15,16,17,18,19,20,21,22,23,24,25,26. Hence, various memory devices have been suggested for artificial synapses, which can be classified into two categories: electronic20,25,27,28,29,30 and optoelectronic synapses26,31,32.

In particular, the optoelectronic synapses can detect and memorize optical signals simultaneously, consequently enabling the implementation of a human vision-neural system and an optogenetic neural network system. Thus, optoelectronic synapse devices have attracted considerable research attention20,23,33. During the initial stage of the relevant researches, conventional photodetectors were integrated with nonvolatile memory devices. Wang et al.23 connected an organic heterojunction photodetector to a ferroelectric synaptic device. A light signal received by the photodetector was converted into an electrical signal; subsequently, the electrical signal updated the weight of the synaptic device. Seo et al.20 reported an optic-neural synaptic device for colored and color-mixed pattern recognition by fabricating an optical-sensing device and synaptic device on the same h-BN/WSe2 heterostructure platform. In recent years, optoelectronic synapses have been fabricated by embedding the memory function into photodetector devices using (i) trapping states based on intrinsic defects31 and (ii) heterojunction structures consisting of wide/narrow bandgap materials27,32. Lee et al.31 reported the long-term potentiation (LTP) phenomenon updated by optical spiking signals in an amorphous oxide semiconductor device with the inherent persistent photoconductivity characteristic originating from oxygen vacancies. Pradhan et al.27 implemented an optoelectronic synapse by endowing a memory function to a graphene photodetector using perovskite quantum dots grown in a graphene lattice; they then presented the LTP characteristic by optical spikes and the long-term depression (LTD) characteristic by negative voltage spikes. Similar works for an artificial synapse exploiting organic and inorganic perovskites have also been reported34,35,36. Qian et al.32 demonstrated the LTP characteristic with optical spikes and the LTD characteristic by voltage spikes in an organic heterojunction structure composed of wide/narrow-bandgap materials. In summary, most studies on the optoelectronic synapses have only demonstrated the LTP characteristic of channel conductance by optical spikes. Although some research has presented LTD phenomenon of optoelectronic synapse, this was possible with the assistance of electrical signals. Very recently, few studies on optoelectronic synapses have shown both long-term potentiation and depression (LTP/D) characteristics of channel conductance solely through optical spikes. Ahmed et al.37 demonstrated a fully light-controlled synaptic device using layered black phosphorus (BP), which was potentiated by a 365-nm light pulse and depressed by a 280-nm light pulse. For the practical implementation of the human vision-neural system and optogenetic neural network system with competitive recognition rates, it is very important to achieve the potentiation and depression of synaptic weights using a single wavelength light.

This study reports on a MoS2/h-BN heterojunction-based photoelectroactive synapse operating with co-stimuli consisting of optical and electrical spikes. As a result of applying the optical memory phenomenon in h-BN38 to a synaptic device, the synaptic device exhibited a unique property which showed that weight may be potentiated and depressed optically, where the weight-updating modes are determined by the condition of the electrical spike. This bidirectional weight-update operation physically originates from the ionization and neutralization of inherent defects localized in the h-BN via optical pulses under negative and positive gate-bias conditions, respectively. First, the weight-update mechanism is verified via photoluminescence (PL) and Kelvin probe force microscopy (KPFM) analyses. Second, various electrical measurements are used to investigate (i) analog memory characteristics, such as reconfigurability and nonvolatility of multilevel conductance states, and (ii) dynamic characteristics, such as excitatory postsynaptic current (EPSC) and LTP/D. Finally, the training/inference tasks using electrocardiography (ECG) signal data are conducted to confirm that the proposed synapse device can be extended for HNNs.

Results

Light and voltage co-stimulated synapse based on MoS2/h-BN heterostructure

Synapses that exist between neurons modulate their synaptic weights by the timing difference between presynaptic and postsynaptic signals; they also memorize the weights over the short- or long-term periods39,40,41. To mimic such biological synapse functions, we have fabricated a synaptic device operating via light and voltage co-stimuli on the MoS2/h-BN heterojunction structure, as shown in Fig. 1a. In this synaptic device, h-BN with photoactive-inherent defects was used to memorize the weights, and an MoS2 channel with significantly few surface defects was adopted to form a high-quality interface with the h-BN, which facilitated maximization of the weight control via the h-BN. This was possible because it is easy not only to form a heterojunction due to the very low surface defect density of the van der Waals (vdW) materials but also to implement various memory characteristics owing to the diverse bandgap properties of the vdW materials. A highly doped Si electrode served as a weight-control terminal, and two metal electrodes on the MoS2 structure were used as the presynaptic and postsynaptic terminals. In particular, a synaptic weight represented the conductance of MoS2 between the pre/postsynaptic terminals, which was potentiated or depressed by using the co-stimuli of optical and electrical signals. Here, optical spikes generate carriers from the inherent photoelectroactive defects in the h-BN dielectric layer, and voltage spikes transfer the photogenerated electrons or holes into the MoS2 region according to the direction of gate electric field, consequently inducing a nonvolatile change in the MoS2 conductance38. The device configuration, including the constituent materials and dimensions, was verified through various avenues, such as optical microscopy analysis, atomic force microscopy analysis, and Raman spectroscopy measurements (see Supplementary Fig. 1).

Fig. 1: Light and voltage co-stimulated synapse based on MoS2/h-BN heterostructure.
figure 1

a Schematic illustration of device structure and bidirectional weight operation process. b Postsynaptic current (Ipost) responses to the co-stimuli consisting of optical and electrical spikes. c Energy band diagrams indicating weight-updating principles of long-term potentiation and depression (LTP/D). d Photoluminescence (PL) spectra of h-BN. e Kelvin probe force microscope (KPFM) work function mapping image of h-BN and MoS2 before and after ionization of h-BN defects.

Full size image

Next, the synaptic device’s postsynaptic current (Ipost) responses to the co-stimulus of optical and electrical spikes were examined. Here, the wavelength (λ), power (Plaser), and width (tlaser) of the optical spike pulse were fixed at 365 nm, 8.1 mW cm−2, and 0.1 s, respectively. As shown in Fig. 1b, the Ipost under Vread of −15 V increased gradually after optical spikes were consecutively applied under the negative write voltage (Vwrite) of −16 V and returned to the initial current value after the “optical spikes” were applied under a positive Vwrite of +30 V. This bidirectional weight update is an extraordinary property because the optical stimulus has generally been employed only as an excitatory spike pulse of synaptic devices. This updating principle can be explained by the ionization and neutralization of defects in the h-BN by the co-stimulus of the optical and electrical spikes (see details in Fig. 1c and Supplementary Fig. 2). When the optical spikes are applied under a negative weight control voltage (see MoS2/h-BN junction in Supplementary Fig. 2a), the electrons are excited from the h-BN’s inherent defects and then transferred to the MoS2. Such ionized defects, which have lost electrons, act as positive fixed charges in h-BN and increase the MoS2 surface potential as well as the electron injection rate from the presynaptic electrode to the channel region. This increases the channel conductance (Gchannel ↑), which is called LTP. In contrast, when optical spikes are applied under a positive weight control voltage (see Supplementary Fig. 2b for details), the electrons photogenerated from the valance band of the h-BN fill the empty defect states in the h-BN. The simultaneously generated holes move to the MoS2 channel region and then recombine with electrons in the conduction band of the MoS2. The filled defect states are neutralized, thereby causing the MoS2 surface potential to gradually return to its initial value and eventually reducing the electron injection from the presynaptic electrode. As a result, this channel current suppression decreases the channel conductance (Gchannel ↓), which is called LTD. Various analyses based on PL, ultraviolet-visible spectroscopy (UV–Vis), and KPFM measurements have supported this photoelectroactive synaptic operation. As shown in Fig. 1d, the PL analysis verified two distinguishable energy states with the binding energies of 2.2 and 2.8 eV inside the h-BN, which are known to originate, respectively, from a nitrogen vacancy occurring when a nitrogen atom moves to a boron vacancy and a natural nitrogen vacancy formed during material synthesis (see the insets in Fig. 1d and Supplementary Fig. 3b)42,43. The UV–Vis result showed that the energy bandgap of the h-BN was ~4.9 eV, which was much higher than the photon energy of the optical 365-nm laser source (3.4 eV) (Supplementary Fig. 4). This indicates that the nonvolatile conductance change originated from the ionization and neutralization of defects by the optical spikes. The KPFM analysis showed that the work function of the h-BN decreased from 4.91 eV to 4.65 eV and that of the MoS2 decreased from 4.74 eV to 4.71 eV after optical spikes were applied under the negative Vwrite (Fig. 1e). This clearly indicates that the co-stimulus of optical and electrical spikes ionized the inherent photoelectroactive defects in h-BN, and the electrons transferred to MoS2 contributed to increasing the surface potential. Such nonvolatile synaptic plasticities were observed only when optical and electrical spikes were applied simultaneously (Supplementary Fig. 5). This is probably because it is difficult to ionize the defects in h-BN without optical stimulus, which results in only a volatile current response to the electrical spike.

Multistate memory characteristics of vdW photoelectroactive synapse

To ensure stable performance of the complex neural network system, the synaptic devices’ conductance states that form the neural networks should be accessible at any time and be retainable for a long period of time. Thus, as shown in Fig. 2a-b, the conductance modulation was first evaluated based on the threshold voltage (Vth) shift characteristic controlled by the co-stimulus of the optical and electrical spikes. The Ipost-Vread characteristic curves were also measured under the double-sweep mode of the read voltage (Vread) ranging from −30 to +30 V (forward sweep) and from +30 to −30 V (reverse sweep), where the presynaptic voltage (Vpre) was fixed at +1 V. The wavelength, power, and width of the optical spike pulse were fixed at 405 nm, 65 mW cm−2, and 5 s for negative shift (or 90 s for positive shift), respectively. During the application of an optical spike with negative write voltages (Vwrite) varying from −10 to −30 V, Vth shifted from −2.1 to −23.9 V in the negative voltage direction. By contrast, Vth was modulated from −23.9 to −1.5 V in the positive voltage direction during simultaneous application of the optical spike and positive Vwrite ranging from +10 to +30 V. In addition, we investigated the dependency of the Vth shift on the light wavelength and confirmed that higher photon energy enhanced the neutralization process (Supplementary Fig. 6). This Vth modulation was observed only in the devices including the h-BN layer (Supplementary Fig. 7), which supports the physical mechanism that the photoelectroactive synaptic operation was based on the ionization and neutralization of inherent defects in the h-BN. As quantitative results, the densities of the interfacial defects of h-BN/MoS2 and the photoelectroactive defects in h-BN are found to be 4.36 × 1010 cm−2 and 3.88 × 1012 cm−2, respectively (Supplementary Fig. 8). During the ionization and neutralization of photoelectroactive defects, the gate leakage current was sufficiently blocked by the dielectric layers of SiO2 and h-BN (Supplementary Fig. 9). We then examined the cycle-to-cycle variations in Vth values with the optical pulses under Vwrite values of +30, −10, −20, and −30 V. Here, the Vth in each cycle was modulated sequentially by Vwrite from +30 to −30 V, and the pulse width of the optical spike (tlaser) was 60 s. As shown in Fig. 2c, the four different Vth values presented similar values in all cycles, where the relative standard deviation (RSD) was 8.6% for the Vwrite of +30 V and reached 0.4% for the Vwrite of −30 V. This result shows that the Vth modulation method is reliable, and the specific conductance state is therefore accessible at any time. Furthermore, the synaptic device used in this study exhibited outstanding retention characteristics (Fig. 2d). The Vth value obtained via the co-stimulus of the optical (405 nm, 65 mW cm−2, and 5 s) and electrical (−30 V) spikes were maintained even after 21 days, with an RSD of only 1.3%. This is an especially important characteristic in terms of implementing HNNs in that the trained complex information can be retained for a long time. Next, the synaptic device’s current response during application of weaker co-stimuli consisting of optical (405 nm, 65 mW cm−2, and 5 s) and electrical (−20 V) spikes consecutively five times was assessed. As shown in the inset and main graphs of Fig. 2e, the Ipost–Vread characteristic curve was slightly shifted in the negative direction; simultaneously, the Ipost values at the Vread range between −17 and −13 V increased gradually. This is because the inherent defects in the h-BN were ionized even by the weaker co-stimuli, which gradually increased the electron injection probability from the presynaptic electrode to the MoS2. This small and continuous change in Ipost can also be interpreted as a gradual modulation of channel conductance (G), where G is extracted by dividing Ipost by Vpre (G = Ipost/Vpre). As shown in Fig. 2f, the G value at the Vread of −15 V increased gradually from 0.15 to 0.30 μS when co-stimuli were applied consecutively to the device. This trend was observed for all Vread cases, which indicated that the synaptic device’s conductance states can be further subdivided using a weaker co-stimulus with a smaller pulse width (for optical spikes) or amplitude (for electrical spikes). In addition, such a conductance modulation range can be adjusted by selecting an appropriate Vread to ensure the corresponding neural networks’ optimal performance. These conductance modulation characteristics were investigated further, as described in the next section.

Fig. 2: Multistate memory characteristics of vdW photoelectroactive synapse.
figure 2

a Negative threshold voltage (Vth) shift characteristics controlled by optical spikes under negative voltage bias. b Positive Vth shift characteristics controlled by optical spikes under positive voltage bias. c Cycle-to-cycle variations in Vth with the optical spikes under various Vwrite values. d Retention characteristics of the modulated Vth. e Negative Vth shift in response to five consecutive co-stimuli spikes. f Interpretation of (e) in terms of the synaptic weight modulation with respect to the number of co-stimuli. All Vth values were extracted from positive gate sweep data.

Full size image

Dynamic characteristics of vdW photoelectroactive synapse

Artificial synapses constituting HNNs need to modulate their synaptic weight persistently in response to continuous stimulus signals. For further investigation of such synaptic dynamics, we analyzed the change in the channel conductance (G) that occurred when co-stimuli consisting of various optical and electrical spikes were applied. Here, all measurements for G were conducted under a presynaptic voltage (Vpre) of 1 V and a read voltage (Vread) of −15 V. As shown in Fig. 3a, when 10 optical spikes of 365 nm, 2.1 mW cm−2, and 0.2 s were applied to the device under electrical spikes of −16 V, G increased gradually from 0.03 to 13.1 nS. Subsequently, it quickly decreased to ~1.1 nS and its value was maintained even after 400 s. This conductance response may have been a result of two major factors: (i) the photogenerated electron-hole pairs in the MoS2 channel layer that disappeared quickly and (ii) the positively ionized inherent defects in the h-BN dielectric layer that maintained for a long period of time31,38,44. It appears that the former effect completely disappeared after ~270 s, and consequently, only the conductance difference (ΔG) resulting from the latter was maintained persistently. This trend was observed similarly in all cases when many co-stimuli were applied (lower panel of Fig. 3a). As the number of co-stimuli increased from 10 to 50, the ΔG values immediately after removal of the co-stimuli and after retention for 400 s gradually increased from 13.1 to 267.1 nS and from 1.1 to 57.2 nS, respectively.

Fig. 3: Dynamic characteristics of vdW photoelectroactive synapse.
figure 3

a Conductance (G) response decaying characteristic (upper panel) and retained conductance difference (ΔG) immediately after application of the co-stimuli and after 400 s (lower panel). b Dynamic range of conductance (Gmax/Gmin) with respect to light power. c Extracted dynamic range with respect to the variation of light pulse width and voltage amplitude. d Long-term potentiation/depression (LTP/D) characteristics of the photoelectroactive synaptic device. e Dynamic range (upper) and asymmetricity (lower) vs. the number of pulses for the LTP/D curves. f Conductance difference between n-th and (n + 1)-th states (ΔG) vs. pulse number (upper) and standard deviation (SD) of ΔG vs. pulse number (lower).

Full size image

We then investigated the dynamic range of the conductance (Gmax/Gmin) modulated by 100 co-stimuli consisting of various optical (Variation cases #1, #2) and electrical (Variation case #3) spikes. The dynamic ranges for the three variation cases are displayed in Fig. 3b, c. Here, the power of the optical spike (Plaser), the pulse width of the optical spike (tlaser), and the pulse amplitude of the electrical spike (Vwrite) varied from 2.24 to 17.52 mW cm−2, from 0.1 to 0.4 s, and from −15 to −18 V, respectively. The dynamic range values increased for all three variation cases, which is attributed to the increment of ionized inherent defects in the h-BN when increasing Plaser, tlaser, and Vwrite of the co-stimuli. The synaptic device used in this study exhibited an excellent Gmax/Gmin characteristic, and the highest Gmax/Gmin of 4363 was achieved when the tlaser and Vwrite were 0.2 s and −18 V, respectively. This result is significant as a high dynamic range in conductance modulation is one of the desirable requirements to secure a high pattern recognition rate of the HNN45,46. We also investigated paired-pulse facilitation (PPF) characteristics. The obtained PPF ratio was 239% for ∆t = 0.5 s which exponentially decreased to 82% for ∆t = 5 s (Supplementary Fig. 10).

In addition, this synaptic device exhibited excellent LTP/D characteristics during the application of 200 continuous co-stimuli composed of optical and electrical spikes. As shown in Fig. 3d, the G value increased gradually from 0.12 to 44.9 nS when optical spikes were consecutively applied at a Vwrite of −16 V, and it decreased to the initial conductance value when the optical spikes were applied at +30 V. As the number of applied co-stimuli increased from 50 to 200, Gmax/Gmin increased significantly from 153 to 785. With a tradeoff arising between the dynamic range and linearity (or symmetricity), the asymmetricity (AS) also increased from 3.22 to 5.71 (Fig. 3e), but the increased AS value was still acceptable for application to HNNs. Here, the AS value was calculated by subtracting the nonlinearity value of the LTD curve from that of the LTP curve (AS = NLP − NLD). A detailed explanation of the nonlinearity extraction is shown in Supplementary Fig. 11. Such a tradeoff can be explained by investigating the conductance change (ΔG) and standard deviation (SDΔG) observed in the LTP/D characteristics (Fig. 3f and Supplementary Fig. 12a–b). A large ΔG results in a high Gmax/Gmin, and a large SDΔG degrades the NL values of the LTP/D characteristics. As shown in the upper panel of Fig. 3f, the ΔG by a unit co-stimulus increased gradually as the pulse number increased. Accordingly, the Gmax/Gmin after the application of 100 co-stimuli also increased considerably. The SDΔG also increased from 0.39 to 0.93 nS as the applied co-stimuli increased from 50 to 200, thus degrading the linearity and resulting in a gradual increase in the AS value (lower panel of Fig. 3f). In addition, we extracted the effective number of conductance states between Gmax and Gmin (NSeff) when the ratio of ΔG to Gmax−Gmin (Δ/Δmax) was >1, 0.5, and 0.1% (Supplementary Fig. 12c–d). For all pulse-number cases, the NSeff value increased significantly as the Δ/Δmax decreased from 1 to 0.1%. Especially, in the 100-pulse case, the NSeff in the LTP and LTD regions reached 91 and 81, respectively, even under the strict condition: Δ/Δmax > 0.1%. This means that 91 and 81% of the total conductance states are effectively usable, indicating that our photoelectroactive synapse possesses enough number of conductance states between Gmax and Gmin. We also confirmed that our synaptic device responded well even at the pulse frequency of 50 Hz, where the pulse widths for reading and writing were both of 10 ms (Supplementary Fig. 13). The synaptic device exhibited no significant degradation in conductance levels for 200 cycles of potentiation and depression (Supplementary Fig. 14). We also confirmed that the potentiated and depressed conductance levels were retained stably for 24 h by monitoring at an interval of 3 h (Supplementary Fig. 15). Furthermore, a performance benchmark of such synaptic dynamics in comparison with those investigated in other optoelectronic synapse studies was conducted (Supplementary Table 1). Compared to other devices, our synaptic device exhibited a notably high Gmax/Gmin (>384) while maintaining a modest AS value (<3.9) and a much longer retention time (>21 days).

Electrocardiogram pattern training and inference task

For the practical application of HNNs in various fields, it is important to test the feasibility of HNNs using various practical data patterns in addition to standard data patterns (e.g., MNIST and CIFAR-10). A training and inference simulation using electrocardiogram (ECG) patterns was then conducted to confirm the feasibility of the vdW photoelectroactive synapse-based HNNs. As the ECG waveforms are very similar to each other (Fig. 4a), it becomes difficult to classify them without medical expertise, which creates limitations for application to smart healthcare systems47. In this technical circumstance, HNNs may help in performing an accurate diagnosis because they can learn various patterns with ambiguous differences effectively and distinguish such small differences probabilistically. According to various abnormal symptoms in the human heart, the ECG waveforms can be divided into five types––N, S, V, F, and Q––whose annotations are summarized in Supplementary Table 2. To prepare the ECG data patterns for training and testing, a single ECG waveform consisting of five specific waves was first extracted from continuous waveforms (see Supplementary Fig. 16 for details), and its electrical potential values were normalized from −1 to 1 (step #1: preprocessing). The normalized waveforms were then sampled every 8 ms (total of 160 discrete time points) (step 2: sampling). As a result, the ECG patterns consisting of 160 sampling data points were prepared for training and testing. For the training and inference simulation, a single-layer artificial neural network (ANN) was theoretically constructed with a size of 160 × 5 using nonlinear synaptic characteristics such as Gmax/Gmin and AS, where other factors including cycle-to-cycle and device-to-device variations remained ideal (see Supplementary Fig. 17 for details). The ANN consisted of 160 input neurons, 5 output neurons, and 160 × 5 synapses, which fully connected them (Fig. 4b). The detailed learning process is as follows. The prepared ECG patterns were applied to the input layer in the form of an input voltage vector (V) comprising 160 input voltages (V1–160) and transformed to the output vector (y) through the matrix–vector product with a synapse matrix (WV) and an activation function (y = f(WV)). Here, the synaptic weight (Wn,m) can be expressed as the difference between the conductance values of two equivalent photoelectroactive synapses, Wn,m = Gn,m+ − Gn,m− (Fig. 4c). This is because the conductance of artificial synapses implemented in hardware is always positive, unlike that of the synapses in software, which have both positive and negative values. Then, the output vector was compared to the label vector of the output neuron (k) to determine the sign of the weight update (sgn(∆W)), and all synaptic weights in the synaptic matrix were modulated according to the determined sign (see the METHODS section for details). This process was repeated multiple times to train 1500 training ECG patterns to the designed ANN. Then, the training accuracy was calculated using 600 separate testing patterns for each training cycle. The mapping images of the synaptic weight matrix after the 10th and 1000th epochs are displayed in Fig. 4d, where the 160 synaptic weights for each ECG type are arranged in rows in each column. After the 10th training epoch, unclear mapping images are observed, however, after the 1000th epoch, the synaptic weight values of a specific area for a certain ECG type have been updated to be clearly distinguishable. The confusion matrix between the actual (input) and predicted (output) ECG types also shows that the designed ANN was well-trained after the 1000th epoch (Fig. 4e). The sensitivity (S) for each type was high overall (>90%), and the maximum sensitivity reached 95.8% in type “N.” Here, the sensitivity for each type (Sn) was calculated as the percentage of the predicted outputs (On) over the total actual inputs (In) of the corresponding type (Sn = On/In). Finally, the training accuracy was investigated with respect to the number of pulses ranging from 50 to 100 (Fig. 4f). For all pulse numbers, the accuracy increased according to the training epoch, and the maximum accuracy of 96.1% was achieved for 150 pulses. This ECG pattern recognition task is meaningful in that the pattern recognition technique via ANN can be applied to diagnose not only cardiac arrhythmia but also various diseases. In addition, a multilayer ANN simulation was conducted for MNIST digit patterns using the platform “MLP+NeuroSim ver. 1.0.” The maximum accuracy achieved was 91.1% for 100 pulses (96.2% for the ideal case), indicating the applicability of this synaptic device to HNNs (Supplementary Fig. 18).

Fig. 4: Electrocardiogram pattern training and inference task.
figure 4

a Five types of electrocardiogram (ECG) waveforms and preparation process for training and testing datasets. b Schematic illustration of the designed single-layer neural network with a size of 160 × 5. c Representation of synaptic weight using the conductance difference between two equivalent photoelectroactive synaptic devices. d Mapping images of 800 synaptic weights after 10th and 1000th epochs. e Confusion matrix of heartbeat type classification for 600 test datasets. f Training accuracy as a function of the number of training epochs according to the number of pulses.

Full size image

Discussion

In this study, a photoelectroactive synapse based on the MoS2/h-BN heterojunction was implemented, where the synaptic weight was potentiated and depressed via “optical spikes” under negative and positive gate-bias conditions, respectively. Through PL, UV–Vis, Raman, and KPFM analyses, it was verified that such a unique bidirectional weight-update operation originated from the inherent defects in the h-BN. The defects were ionized and neutralized via co-stimuli comprising optical and electrical spikes. The photoelectroactive synapse exhibited high accessibility (cycle-to-cycle variation of <1%) and outstanding retention (>21 days) characteristics. Furthermore, important synaptic dynamics, such as EPSC and LTP/D, were demonstrated by continuous co-stimuli, and high Gmax/Gmin (>384) and modest AS (<3.9) values compared to those of other optoelectronic synaptic devices were successfully achieved. Finally, the feasibility of the proposed synaptic device’s application to HNNs was examined using ECG pattern training and inference tasks. A very high accuracy of 96.1% was achieved in a simulated single-layer ANN with a size of 160 × 5. Overall, we believe that this unique bidirectional weight update concept will play an important role in the implementation of optoelectronic HNNs in the future. We also note that, for the mass production of 2D heterostructure devices including this optoelectronic synapse, there are still several technical issues to be solved in 2D material processing technology: (i) large-scale growth of various 2D materials including dielectric and semiconducting materials and (ii) direct growth on the surface of 2D materials for the formation of heterojunction structures. Some pioneer works have also been reported recently for the mass production technology of 2D materials such as wafer-scale CVD growth, wafer-scale transferring, direct epitaxial growth on the 2D material surface48,49,50,51,52,53, but further technological advances like doping and contact engineering are still required.

Methods

Fabrication of photoelectroactive synaptic devices

A heavily p-doped silicon substrate with a thermally grown 90-nm-thick SiO2 gate dielectric was used as the bottom gate. The h-BN flakes were mechanically transferred onto the SiO2 via a residue-free transfer method based on adhesion energy engineering20. Then, the MoS2 flakes were transferred onto the h-BN via the same transfer method. The postsynaptic and presynaptic electrodes were patterned on the MoS2/h-BN structure using electron-beam lithography, followed by 10-nm-thick Ti and 50-nm-thick Au electrode deposition. Here, the distance and width between the two electrodes were 2 and 8 μm, respectively.

Characterization of the photoelectroactive synaptic devices

Top-view optical images of the devices were obtained using an upright metallurgical microscope (Olympus BX53M). Electrical measurements of the synaptic devices were recorded using a Keithley 4200A-SCS parameter analyzer. A constant voltage of 1 V was applied to the presynaptic terminal (Vpre) to read the postsynaptic current (Ipost) flowing between the pre- and postsynaptic terminals. For the application of the co-stimuli, a UV laser source with a wavelength of 365 nm was connected to the parameter analyzer and then used on the synaptic device simultaneously with the electrical voltage spikes.

Steady-state PL measurements at 70 K were performed using a LabRAM HR Evolution Visible/NIR (Horiba Jobin Yvon, France) with a 325-nm argon laser excitation source (a spot size of 1–2 µm) and a charge-coupled device (CCD) detector. In addition, absorption measurements were conducted using a UV-3600 Shimadzu UV-Vis-NIR spectrophotometer at wavelengths between 200 and 3000 nm at room temperature. Here, a deuterium and halogen lamp excitation source was used along with a photomultiplier tube, InGaAs, and PbS detector.

AFM and KPFM measurements were performed in the dark and under ambient conditions using an NX10 system (Park Systems Corp.). For the KPFM measurement, VG was applied to the back gate, and the source/drain electrodes were grounded. A platinum/iridium (Pt/Ir)-coated Si tip was used, and it was calibrated on a highly oriented pyrolytic graphite (HOPG) surface. The surface work function of the samples was obtained from the contact potential difference (CPD) between the work functions of the tip and HOPG (Ftip−FHOPG = VCPD), where a standard HOPG value of 4.6 eV was considered.

Weight update for photoelectroactive synaptic devices

According to the determined sign of the weight update, synaptic devices were potentiated or depressed according to the following equations:

$$sgn(\Delta {{{\mathbf{W}}}}) = + 1 \to G_{{{{\mathrm{n}}}} + 1} = G_{{{\mathrm{n}}}} + \Delta G_{{{\mathrm{P}}}} = G_{{{\mathrm{n}}}} + \alpha _{{{\mathrm{P}}}}e^{ - NL_{{{\mathrm{P}}}}\frac{{G_{{{\mathrm{n}}}} - G_{{{{\mathrm{min}}}}}}}{{G_{{{{\mathrm{max}}}}} - G_{{{{\mathrm{min}}}}}}}}\left( {\Delta G \,>\, 0,\;{{{\mathrm{potentiated}}}}} \right)$$
(1)
$$sgn(\Delta {{{\mathbf{W}}}}) = - 1 \to G_{{{{\mathrm{n}}}} + 1} = G_{{{\mathrm{n}}}} + \Delta G_{{{\mathrm{D}}}} = G_{{{\mathrm{n}}}} - \alpha _{{{\mathrm{D}}}}e^{NL_{{{\mathrm{D}}}}\frac{{G_{{{{\mathrm{max}}}}} - G_{{{\mathrm{n}}}}}}{{G_{{{{\mathrm{max}}}}} - G_{{{{\mathrm{min}}}}}}}}\left( {\Delta G \,<\, 0,\;{{{\mathrm{depressed}}}}} \right)$$
(2)

Here, Gn+1 and Gn denote the synaptic conductance values when the n+1th and nth spikes were applied, and the parameters α and NL indicate the magnitude of conductance change and the nonlinearity, respectively.