An Image Detection–Memory–Recognition Artificial Visual Unit Based on Dual‐Gate Phototransistors

Optoelectronic synapses integrating sensing–memory–processing functions have great advantages in neuromorphic computing for visual information processing and complex learning, recognition, and memory in an energy‐efficient manner. Herein, a light‐induced bidirectional response is demonstrated in the proposed WSe2/MoS2 junction field‐effect transistor (JFET) with an extra Ge back gate. The WSe2/MoS2 JFET exhibits high responsivity and detectivity owing to effective modulation by the top junction and back dielectric gates. The unique bidirectional photoresponse and interfacial state storage properties of the device result in significant synaptic excitatory effects under visible light stimulation and remarkable synaptic inhibitory effects under infrared illumination. Optical storage with visible light and optical erasure with infrared light can be achieved in the device based on its synaptic behavior. The conductance changes under different types of illumination can be used to mimic the weight update in a neural network for image recognition. A high accuracy exceeding 98% was achieved in handwritten digit recognition. This visible and infrared dual‐band phototransistor, which integrates image perception, memory, and recognition functionalities, provides a promising solution for self‐driving, surveillance, computer vision, and biomedical imaging applications.

neuromorphic visual systems cannot respond to optical stimuli directly. They require image sensor arrays to convert optical signals into electrical signals so that these signals can be passed to neuromorphic chips for further signal processing. [14][15][16][17] In contrast, the sensory neurons in the human visual system can not only detect light stimuli but also perform first-stage image processing before more complex visual signal processing occurs in the visual cortex of the human brain. [18][19][20] Using a photoelectric neural network to simulate artificial visual systems is an effective way to realize image recognition. There is a high demand for the development of multifunctional devices that can integrate sensing, memory, and processing functions in more efficient artificial visual systems.
Optoelectronic synaptic devices are potential devices for artificial visual systems owing to their inherent light sensitivity and memory characteristics. [21][22][23][24][25][26][27] The capability of these devices to perform image processing and neuromorphic computation for machine vision has been demonstrated in recent studies, [13,[28][29][30][31] and some optoelectronic synapses with integrated image sensing capabilities have been designed to simplify the circuitry of neuromorphic visual systems. [32][33][34][35][36][37] However, the inevitable use of a combination of electrical and optical signals or complicated device structures to perform neuromorphic computation in these devices results in similar problems of hardware redundancy, high-power consumption, and computational latency due to the bandwidth connection-density tradeoff. [38] In addition, there are few reports on fully light-tunable neuromorphic synaptic devices, especially for devices that can be modulated by visible and infrared light. Therefore, visible-infrared fully light-tunable neuromorphic visual devices with sensory-memory-recognition functions, which have the potential to achieve high bandwidths and ultralow power consumption, should be studied. [39,40] In this work, we report a neuromorphic visual device with inherent fully light-modulated behavior in a WSe 2 /MoS 2 heterostructure junction field-effect transistor (JFET) with an extra Ge gate. The device exhibits visible and infrared dual-band detection capabilities as well as visible and infrared wavelength selectivity. Under 532 nm illumination, a positive photoresponsivity is realized through the photoconductivity mechanism of MoS 2 . In contrast, upon 1550 nm illumination, the JFET exhibits a negative photoresponsivity behavior, which is attributed to the photogating effect of Ge. Using these unique optoelectronic characteristics, key synaptic functions such as short-term memory (STM), long-term memory (LTM), long-term potentiation (LTP), and long-term depression (LTD) are emulated and achieved through the tunable photoconductance. In addition, programming and erasing functions and neuromorphic image recognition are independently implemented based on the synaptic properties. The devices proposed in this work are expected to enable the realization of a fully light-controlled integrated information sensing-memory-processing system that can execute real-time image detection, in situ image memorization, and recognition in a single device and thereby avoid the energy consumption and time latency caused by data conversion and transmission. These results present a new paradigm for future photonic neuromorphic circuits and artificial vision systems.

Results and Discussion
A WSe 2 /MoS 2 JFET was fabricated on a SiO 2 (300 nm)/p-Ge substrate using a conventional dry transfer method. The details of the fabrication processes can be found in the Experimental Section. Figure 1a shows a schematic of the JFET in which WSe 2 serves as the top gate material and MoS 2 as the channel material. MoS 2 was selected as the channel material based on its high mobility and on/off ratio, whereas WSe 2 was selected as the top gate to reduce the current at a reverse bias (as shown in Figure 2a). Ge was used as the back gate material. Carrier transport in the MoS 2 can be modulated by both the WSe 2 junction gate and the bottom p-Ge dielectric gate. Multilayer graphene was employed to eliminate the contact barrier between the MoS 2 and Au and to reduce the contact resistance between the WSe 2 and Au. An optical image of the JFET is shown in Figure 1b. The atomic force microscopy (AFM) measurements shown in Figure 1c indicate that the thicknesses of WSe 2 and MoS 2 are approximately 30.4 and 25.3 nm, respectively. As shown in Figure 1d, distinct WSe 2 , MoS 2 , and Ge Raman peaks were observed. The two peaks located at approximately 249 and 257 cm À1 correspond to the E 1 2g and A 1g modes of multilayer WSe 2 , respectively, whereas the Ge peak is located at approximately 300 cm À1 . The other two distinct Raman peaks located at approximately 382 and 407 cm À1 correspond to the E 1 2g and A 1g modes of multilayer MoS 2 .
The current-voltage (I-V ) curve of the WSe 2 /MoS 2 heterostructure exhibits a notable rectification behavior (black and red curves in Figure 2a) with a rectification ratio of up to 2.5 Â 10 2 , which confirms the formation of a p-n junction between WSe 2 and MoS 2 . The blue line shows the I-V curve of MoS 2 with floating top (WSe 2 ) and bottom (Ge) gates. It indicates that good conductivity and symmetry are maintained at both ends of MoS 2 . Figure 2b shows the transfer characteristics of the top gate in which an on/off ratio of approximately 10 3 is observed. The device exhibited a minimum subthreshold swing of approximately 95 mV dec À1 , which is superior to that of conventional MoS 2 metal oxide semiconductor field-effect transistors (MOSFETs). Figure 2c shows the transfer curves (drain current-bottom gate voltage, I d -V bg ) of the device at V ds = 1 V when V tg was in the range from À2 to 3 V. Notably, a shift in the threshold voltage (V th ) was observed at different V tg . At negative V tg , a larger depletion region was formed in the WSe 2 /MoS 2 heterojunction; thus, the MoS 2 channel was depleted at a smaller bottom gate voltage, resulting in a positively shifted V th . At positive V tg , the depletion region was reduced, leading to increased I ds ; thus, a more negative V tg was required to deplete the channel, resulting in a negatively shifted V th . This conspicuous tunable mobility of the JFET presents a significant advantage over traditional MOSFETs. The output characteristics (drain current-drain voltage, I d -V d ) at different bottom gate voltages are shown in Figure 2d. The characteristics exhibit the three typical transistor stages of 1) linear, 2) pinch-off, and 3) saturation trends. As the voltage of the top gate (V tg ) increased from negative to positive, the channel was switched ON and the current gradually increased. Figure 2e Figure 3d,e. The responsivity (R), which is defined as R = I ph /P, where I ph is the photocurrent and P is the incident light power, is a key parameter for evaluating the performance of a photodetector. The detectivity (D*), which is defined as D* = R (A/2eI dark ) 1/2 assuming that the noise current is dominated by the dark current, where I dark is the dark current, is another key parameter for characterizing the sensitivity of a photodetector. The photodetection performance of the JFET under the incidence of 532 nm light at an intensity of 1.8 nW is shown in Figure 3d,e. The responsivity and detectivity of the device could be effectively modulated using the dual gates. When V bg was in the range from À4 to 2 V, a high responsivity (R > 50 A W À1 ) could be obtained because of the large photocurrent due to the high channel conductance and carrier mobility. www.advancedsciencenews.com www.advintellsyst.com The highest responsivity of 67.3 A W À1 was obtained at a V tg of À2 V and V bg of À1.5 V. When V bg was in the range from À7 to À3 V, a high detectivity (D* > 5 Â 10 12 Jones) could be obtained because of the high light/dark current ratio near the threshold voltage. The highest detectivity of 1.4 Â 10 13 Jones was obtained at a V tg of À2 V and V bg of À6.3 V. The photocurrent-time (I-t) curves under illumination by 532 nm visible light at the top gate voltages of À1, 0, and 1 V and floating bottom gate are shown in Figure 3f. The light response remained almost unchanged under different gate voltages. The absolute values of the negative response and detectivity under 1550 nm illumination are shown in Figure 3g,h. A higher responsivity (R > 6 mA W À1 ) was obtained when V bg was in the range from À4 to 2 V and V tg in the range from À0.5 to À3 V. The highest responsivity of 10.28 mA W À1 was obtained at a V tg of À1 V and V bg of 0.3 V. A higher detectivity was obtained when V bg was in the range from À6 to À2 V and V tg in the range from À0.5 to À3 V. The highest detectivity of 8.7 Â 10 8 Jones was obtained at a V tg of À1 V and V bg of À5 V. Figure 3i shows the I-t response under 1550 nm light. The light response at a top-gate voltage of À2 V is twice that at a top-gate voltage of 2 V, which is consistent with Figure 3g. To further understand the photoresponse mechanism of the JFET photodetector, a carrier transport model was established ( Figure S2, Supporting Information). Effective modulation of the infrared response was achieved through the tunable effect of the dual gates in the channel depletion region.
Owing to its controllable optoelectronic properties of both positive and negative responses, the optoelectronic device has a strong potential in mimicking different types of neuroplasticity. The specific light wavelengths of 532 and 1550 nm were selected to trigger changes in the synaptic weights through the enhancement and weakening of the synaptic connections to mimic excitatory postsynaptic currents (EPSCs) and inhibitory postsynaptic currents (IPSCs), respectively. The EPSCs of the device triggered by 532 nm light pulses of different frequencies are shown in Figure 4a. A greater degree of excitation occurred at higher frequencies. The device can exhibit STM, LTM, and a transition from STM to LTM. When the input light pulses were of a low frequency (0.1 Hz), the synaptic weight could persist for a few milliseconds before quickly decaying to the original state after removing the optical stimuli, which corresponds to STM in biology. However, as the frequency of the input light pulses increased, the synaptic weight lasted for a longer time after removing the stimuli, which is analogous to the transition from STM to LTM. Furthermore, the devices exhibited light-powertunable synaptic characteristics, as shown in Figure 4b. The EPSC increased with the optical power. Paired pulse facilitation (PPF) is a typical short-term synaptic plasticity triggered by two consecutive presynaptic spikes and is essential for decoding temporal vision information in biological neural systems. When two consecutive light pulses (532 nm, 25 mW cm À2 ) were applied to the device, the second enhanced EPSC (A 2 ) was larger than the first EPSC (A 1 ), as shown in Figure 4c. The extent of enhancement was strongly dependent on the interval between the two consecutive light pulses. Figure 4c shows that the PPF ratio www.advancedsciencenews.com www.advintellsyst.com (A 2 /A 1 ) gradually decreased with the increasing period (pulse width = 100 ms). The IPSC of the device triggered by an optical pulse of 1550 nm wavelength (32 mW mm À2 ) is shown in Figure 4d,f. The amplitude of the IPSC can be modulated by the number of applied 1550 nm spikes. As shown in Figure 4d, as the spike number increased from 30 to 90, the IPSC increased from 6.2 to 12.2 nA. In addition, the change in ΔPSC (the difference between the current corresponding to the last light pulse and the original state current) was also dependent on the pulse frequency and increased with the pulse frequency. As shown in Figure 4e, when the pulse frequency was 5 Hz, the current decreased slowly, whereas when the frequency increased to 10 Hz, the current dropped faster. The rate of current dropped increases with frequency. Similar to 532 nm synaptic stimulation, PPF characteristics were also exhibited during 1550 nm inhibition. Figure 4f shows IPSCs induced by a pair of inhibitory optical pulses (1550 nm, 32 mW mm À2 ) with a pulse width of 50 ms. The peak values for the IPSC triggered by the second optical spike were significantly larger than those triggered by the first optical spike, and the PPF ratio (A 2 /A 1 ) gradually decreased with increasing period (pulse width = 10 ms). The energy consumption of the EPSC and IPSC are shown in Table S1, Supporting Information. Although the light energy consumption is not significantly lower than those reported in other studies, it can be improved by increasing the light frequency.
In addition, to study the synaptic mechanism, we investigated the hysteresis characteristics of the device. The transfer characteristics of the bottom gate transfer curves are shown in Figure S3, Supporting Information. The synaptic characteristics of the device can be attributed to charge trapping and detrapping at the interface regulated by light. Under 532 nm light pulses, the electrons were trapped at the MoS 2 /SiO 2 interface and then slowly released to achieve a device storage effect. Under 1550 nm illumination, the accumulation of electrons at the Ge/SiO 2 interface resulted in the expulsion of electrons at the MoS 2 /SiO 2 interface, which led to a continuous drop in the IPSC.
Long-term synaptic plasticity characteristics such as LTP and LTD of the synaptic weights are crucial for implementing neuromorphic computational functions. Considering the previous observation that EPSC and IPSC characteristics were obtained by introducing light pulses at critical wavelengths, LTP and LTD were, respectively, achieved only by applying 532 and 1550 nm light pulses with certain intensities and repetition intervals. As shown in Figure 5a, EPSCs were obtained with less than 10 light pulses (532 nm, width = 100 ms, interval = 100 ms), and the current decreased slightly when the light was removed. However, the current did not return to its initial state. In other words, a clear memory effect was observed after the light pulse group was repeated for several cycles. The combined effects of LTP and LTD are shown in Figure 5c. A repeatable bidirectional photoresponse of seven cycles was obtained by applying 10 identical pulses of 532 nm light for LTP and 20 identical pulses of 1550 nm light for LTD. Notably, the repeatable LTP and LTD can be regarded as memory and erasure, respectively, and a faster depression occurred compared to the current drop rate without 1550 nm illumination. Based on the above memory and erasure characteristics, the imaging, storage, and erasure processes of images were simulated, as shown in Figure 5b,d. Figure 5b shows the illumination of three different patterns designated as "X", "D", and "U" with 30 consecutive short optical pulses (with a pulse width of 100 ms and power density of 24 mW cm À2 ) at 5 Hz. The current in each pixel was measured after illumination stimulation and 1 min later to evaluate the Figure 5. a) PSC modulation following exposure to 10 repeated light spikes at a wavelength of 532 nm (24 mW cm À2 , width = 100 ms, interval = 100 ms) and without 1550 nm light. b) "X", "D", and "U" patterns encoded by optical pulses of 532 nm wavelength with a power density of 24 mW cm À2 (pulse width = 100 ms and frequency = 5 Hz) and image state after the light signal was removed for 1 min. c) PSC modulation following exposure to 10 repeated light spikes at a wavelength of 532 nm (24 mW cm À2 ) and 20 repeated light spikes at a wavelength of 1550 nm (32 mW mm À2 ) with the same width and interval of 100 ms for emulating the LTP and LTD of synapses, respectively. d) Encoding of "X", "D", and "U" patterns by 532 nm optical pulses with a power density of 24 mW cm À2 (pulse width = 100 ms and frequency = 5 Hz), and pattern after optical erasure by exposure to a 36 s 1550 nm pulse at 32 mW mm À2 (pulse width = 100 ms, frequency = 5 Hz).
www.advancedsciencenews.com www.advintellsyst.com memory retention. For comparison, the currents in all the pixels were scaled between 0 and 1 where 0 corresponds to the minimum measured dark current before exposure and 1 to the maximum photocurrent measured after exposure. After light exposure, the pattern was well recognized, as can be seen from the relatively high conductance contrast between the illuminated pixels and the pixels without optical stimulation. The illuminated pattern was memorized after 1 min. To sequentially image the next figure, 1,550 nm illumination was introduced to erase the memorized image. As shown in Figure 5d, after pattern recognition under 532 nm light, the imaging array was exposed to 1550 nm (32 mW mm À2 , 5 Hz) wavelength light for 36 s to optically erase the pre-existing information. The increase in the conductance and optical erasure of the stored information highlights the capabilities of the device for in-pixel image enhancement and real-time image acquisition based on input visual information.
When the device is operating as an artificial synapse, synaptic weight changes can be mimicked by illumination with optical pulses, which mimic synaptic stimuli, to trigger changes in the conductance of the device. This modulation of the weight change using illuminating optical pulses, that is, LTP under 532 nm illumination and LTD under 1550 nm illumination, was experimentally measured to simulate artificial neural networks. The strong dependence of the LTP/LTD characteristics on the parameters of the optical stimuli (such as pulse frequency) can be used to simulate synaptic plasticity in a neural network. In our artificial vision unit, a neural network was used to implement image recognition. Five hundred thirty-two and One Thousand Five Hundred Fifty nanometers of optical pulses of different pulse frequencies (1, 2, 3, and 10 Hz) and the respective power densities of 24 and 32 mW mm À2 were used to obtain LTP/LTD characteristics, as shown in Figure 6a. The comparison of the LTP/LTD profiles at different pulse frequencies in Figure 6b (extracted from Figure 6a) clearly shows that the frequency of the optical pulses affected the conductance. A higher optical pulse frequency induced a relatively large weight change.
The above analysis results suggest that the device can be used for imaging and memory functions, as shown in Figure 7aI. In this neuromorphic visual system, visual information is first detected through real-time imaging, and image memory is subsequently realized through the storage array. The stored images are transported to an artificial neural network to complete the image training and recognition functions, as shown in Figure 7a. The device array could be realized later by using large-scale material transfer, e-beam lithography (EBL), e-beam evaporation, and other technologies. Furthermore, owing to the LTP/LTD characteristics of the device under different pulsed-light modulations, weight changes can be effectively achieved in the neural network. Based on the parameters extracted from the conductance update characteristics of the device, the supervised learning of handwritten digits (28 Â 28 pixels) from the Modified National Institute of Standards and Technology (MNIST) database via backpropagation in a threelayer optoelectronic neural network (ONN) with one hidden layer was simulated (Figure 7aII). The conductances of the devices (Gm,n, Figure 7b) were used to modify the synaptic weights of the network. As shown in Figure 7b, the synaptic weight of the network is defined as the conductance difference between two equivalent optical synapses, that is, W = G þ ÀG À , where G þ and G À represent the conductances of the two devices. Using this approach, 60 000 images from the MNIST dataset were trained in our network, and the recognition accuracy for the 10 digits was calculated using the 10 000 test images provided during network training. Figure 7c shows the evolution of the handwritten digit recognition accuracy with the number of training epochs. The calculated recognition accuracy for different pulse widths (1, 2, 3, and 10 Hz) of visible and infrared optical stimuli was used to quantify the synaptic plasticity. Maximum recognition accuracy of 98% was achieved in the optical LTP/LTD pulse simulations. The linearities at different optical pulse frequencies were used to analyze the effect of the frequency on the learning efficiency of the ONN. As shown in Figure 6, there is good linearity in the conductance changes at different frequencies; therefore, the accuracy at different frequencies is relatively high. The specific nonlinear coefficients are shown in Table S2, Supporting Information. Additionally, the conductance during 532 nm stimulation and spontaneous current decrease without 1,550 nm stimulation were used to adjust the weights, as shown in Figure 7c (gray curve). The weight update of the neural network through the inhibitory effect of infrared light on the current effectively improved the accuracy of the neural network from 62% to 98%. www.advancedsciencenews.com www.advintellsyst.com

Conclusion
In summary, a dual-gate phototransistor based on a WSe 2 /MoS 2 JFET with an extra gate was designed. The device exhibited fully light-tunable visible-infrared photoelectric characteristics. Programming and erasing functions and neuromorphic image recognition were implemented in the device. Positive and negative responses were induced by illumination at 532 and 1,550 nm, respectively. The obvious positive response mainly originated from the absorption of visible light by MoS 2 while the negative response was attributed to the photogating effect of the Ge back gate. The bidirectional photoresponses can be utilized to mimic excitatory and inhibitory postsynaptic behaviors, which also enable storage and erasure. The innate optical sensing and memory properties were considered for in-pixel image preprocessing. The device was also deployed in an all-optically driven neural network and the digital image recognition accuracy was increased from 62% to 98%. These results demonstrate that our device provides an attractive visual unit for integrating dualband image perception, memory, and recognition functionalities in neuromorphic computational networks and autonomous vision systems.

Experimental Section
Device Fabrication: A WSe 2 /MoS 2 heterojunction JFET was fabricated on a Ge substrate. A 300 nm thick SiO 2 layer was first deposited. Next, Ni/Au (5/100 nm) was deposited as a metal electrode using e-beam evaporation.
A MoS 2 flake was mechanically exfoliated from bulk single crystals using adhesive tape and transferred onto a patterned substrate using a polydimethylsiloxane (PDMS) stamp. Subsequently, a WSe 2 flake was prepared using the same method and transferred onto MoS 2 under an optical microscope. Subsequently, two more graphene flakes were transferred to the drain and source terminals of the JFET to reduce the contact resistance. Finally, a Gr flake was transferred onto WSe 2 to form the top-gate electrode.
Characterization: The electronic and photoelectronic characteristics of the device were measured using a Keithley 4200A-SCS semiconductor parameter analyzer under dark conditions and under visible and infrared light illumination during the light current (I light ) versus voltage (V ) measurements. AFM (Asylum Research Cipher S) was used to determine the thickness of the 2D materials. The Raman spectra were obtained using a confocal Raman system (WITec α300R) with a 532 nm excitation laser.

Supporting Information
Supporting Information is available from the Wiley Online Library or from the author.