Biological UV Photoreceptors‐Inspired Sn‐Doped Polycrystalline β‐Ga2O3 Optoelectronic Synaptic Phototransistor for Neuromorphic Computing

In this study, the authors fabricate Sn‐doped 100‐nm thick polycrystalline β‐Ga2O3 synaptic field‐effect transistors (FETs) emulating optical and electrical spike stimulation. When stimulated by deep ultraviolet (UV) optical spikes or electric voltage spikes at the gate, the devices exhibit several essential synaptic functions of excitatory‐postsynaptic currents (EPSCs), inhibitory‐postsynaptic currents (IPSCs), paired‐pulse facilitation (PPF), spike‐number‐dependent plasticity (SNDP), and spike‐timing‐dependent plasticity (STDP). Following UV optical stimulation, the devices mimic synaptic plasticity with a photogate effect, and the gate voltage stimulation emulates the synaptic weights according to the state of the gate dielectric interface. The β‐Ga2O3 synaptic FET demonstrates synergistic functions in various optoelectronic stimulation modes and successfully mimics the visual memory formation in bees with UV photoreceptors. Moreover, to verify the translation of optoelectrical‐derived synaptic behaviors of β‐Ga2O3 synaptic FETs into artificial neuromorphic computing, handwritten digit image recognition of the Modified National Institute of Standards and Technology dataset is performed using a convolutional neural network, and a learning accuracy of 96.92% is achieved. The realization of these fundamental functions of biological synapses suggests the utility of Ga2O3‐based optoelectronic devices for next‐generation neuromorphic computing.


Introduction
The human brain is an extremely complex biocomputing system composed of ≈100 billion neurons and 100 trillion synapses capable of simultaneously processing high-level computations consuming only 20 W of power. [1] Recently, artificial intelligence technology that mimics the efficient information processing mechanism of the human brain has emerged as a core technology for big data processing. In particular, a neuromorphic computing chip can directly implement the human brain structure and computational process in hardware that enables efficient lowpower intelligent computation at high speeds. [2][3][4] Furthermore, neuromorphic computing is a promising technology for simultaneously processing and storing data with one component. Consequently, the implementation of neuromorphic computing can overcome the bottlenecks of existing computers based on the von Neumann computing architecture, such as hardware redundancy, power consumption, and computational delay owing to the separation of operating and storage cells. [2,[5][6][7] This enormous potential of neuromorphic computing has prompted a worldwide increase in research to realize the technology. Currently, synaptic devices such as memristors and field-effect transistors (FETs) are studied as hardware neuromorphic computing technology. [8][9][10][11] More recently, complex biological activities and perceptions, such as memory and forgetting processes, classically conditioned learning experiments, and artificial sensory functions, have been emulated using neuromorphic artificial synaptic devices, including memristors and FETs. [8,9] The biological nervous system represents associative learning through sensory systems (vision, hearing, touch, taste, smell, and balance). [12] To comprehensively mimic brain functions, artificial synaptic devices that integrate sensing and processing functions must be developed. Among them, vision plays a vital role in how organisms react to the external environment, primarily because more than 80% of the information acquired through the sensory organs is obtained through vision. [13] In addition to visible light, various animal species contain photoreceptor cells that can absorb UV light in the retina and transduce it into cellular www.advancedsciencenews.com www.advelectronicmat.de signals. [14] For example, bees can be trained to search for sugar water because of their UV-sensitive visual and nervous systems. [15] In the dim Arctic twilight, reindeer use their ability to sense contrast in the UV to locate plants, including lichen and moss, and effectively evade predators (wolves and polar bears). [16] In addition, the UV-specific vision of Pieris rapae plays a crucial role in mate selection. [17] On the other hand, depending on the cumulative exposure amount, intensity, and frequency, UV can cause skin aging, eye damage, and other diseases, in humans. [18,19] Since humans cannot perceive UV light, the development of UV neuromorphic sensors complements our understanding of UV light and can be utilized in a variety of applications, such as chemical/biological sensors, on-chip optical communications, UV astronomy, and early warning systems. [20,21] In particular, the incorporation of light into the operation of synaptic FETs can expand the bandwidth and effectively improve the interconnection problem of synaptic devices. [22] Moreover, the development of synaptic devices with optoelectronic spikes and non-volatile storage functions utilizing synergistic effects between light and electric fields may have great potential for use in hybrid optoelectronic artificial neural networks (ANNs). More specifically, recent research findings have demonstrated non-volatile storage functions in synaptic devices that utilize the persistent photocurrent (PPC) characteristics of oxide semiconductors, such as IGZO, [23][24][25] In 2 O 3 , [26] and ZnO. [27][28][29][30] Synaptic devices of oxide semiconductors are stimulated by deep UV signals owing to their large optical bandgaps. Among them, gallium oxide (Ga 2 O 3 ), which has an ultra-wide bandgap of 4.9 eV is particularly noteworthy. [31] The ultra-wide bandgap provides a controllable number of large memory (storage) states. Moreover, the almost direct bandgap properties of Ga 2 O 3 are suitable for optical device applications and synaptic devices that respond to optical stimuli. Ga 2 O 3 can exist in five crystalline phases: , , , , and . Among them, the phase is mostly used for electronic applications owing to its excellent thermal and chemical stability. [31] Although crystalline Ga 2 O 3 has obvious advantages when used as optoelectronic synaptic devices, most Ga 2 O 3 -based photonic devices have only been studied for photodetector (PD) applications, such as metal-semiconductor-metal PDs, [32][33][34] metal-/Ga 2 O 3 -based Schottky PDs, [35,36] heterojunction PDs, [37][38][39] and phototransistors. [40][41][42] In this study, -Ga 2 O 3 -based synaptic phototransistors were fabricated for the first time. Ga 2 O 3 was deposited via sputtering, which is advantageous for the PPC effect and facilitates uniform film formation over a large area at a low cost. Optical and electrical spikes stimulated the Sn-doped 100-nm thick polycrystalline -Ga 2 O 3 synaptic FETs. The device demonstrates a linear increase in synaptic current in response to UV light stimulation due to the effect of PPC, and potentiation and depression responses by gate voltage stimulation. We show the modulation of weights through the synergistic effect of excitatory and inhibitory synapses and emulate the associative learning process, a representative neurological behavior. Notably, this is the first attempt to artificially simulate the visual memory formation process of organisms with UV photoreceptors, expanding the applicability of Ga 2 O 3 to the neuromorphic sensor field and complementing our understanding of UV light. Moreover, a training method suitable for the conductance behavior of the -Ga 2 O 3 synaptic phototransistors is proposed to further demonstrate the applicability of the synaptic -Ga 2 O 3 phototransistor in the neuromorphic field and a convolutional neural network (CNN) was simulated to perform Modified National Institute of Standards and Technology (MNIST) handwritten digit images pattern recognition.

Results and Discussion
The electrical properties of back-gate Sn-doped 100-nm -Ga 2 O 3 MOSFETs were investigated using the current-voltage (I-V) curves. Figure 1a illustrates the transfer characteristic curves of the drain current (I D ) for a varying gate voltage (V G ) and drain voltage (V D ) measured at room temperature. The Sn-doped 100-nm -Ga 2 O 3 exhibits a gate modulation of ≈10 4 at a drain voltage of 10 V. An output characteristic curve showing typical transistor characteristics consisting of a linear and saturated region is shown in Figure 1b. The increase in I D for the V G step was consistent with the n-type conductivity of the Sn-doped 100-nm -Ga 2 O 3 . The hysteresis of the transfer characteristic curve was observed ( Figure S1, Supporting Information) to investigate the effect of trapping at the Ga 2 O 3 /SiO 2 interface, which is the primary cause of the synaptic behavior. The hysteresis was caused by the trapping and de-trapping of carriers at the interface between the Ga 2 O 3 and SiO 2 gate oxides. This phenomenon is comprehensively demonstrated in the literature for thin-film transistors (TFTs). [23,25,[43][44][45] The charge is trapped only at the shallow level and the hysteresis is relatively small at short sweep delays, whereas the hysteresis becomes larger as the sweep time increases because the charge also gets trapped at the deep level. [46][47][48] To monitor the charge-discharge effect of the traps, the transfer characteristics of the Ga 2 O 3 FET were acquired on "fresh devices" after negative (−30 V) gate stress and positive (+50 V) for 3 s (shown in Figure S2, Supporting Information); the shifts in the threshold voltage (V TH ) confirm that gate stress causes a change in the state of the Ga 2 O 3 /SiO 2 interface.
In biological systems, a synapse forms a junction between the neuronal axon of a presynaptic neuron and the dendrite or cell body of a postsynaptic neuron. The biological synapses and Ga 2 O 3 /SiO 2 -based synapses, which are called synaptic transistors in this study, are displayed in Figure 2d. Information transmission across a synapse begins with the stimulus-evoked release of a neurotransmitter from the synaptic vesicle of a presynaptic neuron into the synaptic cleft. Neurotransmitters are absorbed by the receptors on postsynaptic neurons to generate a postsynaptic current (PSC). Presynaptic pulses can be stimulated by delivering light pulses to the Ga 2 O 3 channel or electrical pulses to the Si back-gate. The carriers in the Sn-doped 100-nm -Ga 2 O 3 channel act as neurotransmitters, and the current collection at the drain becomes equal to that of the PSC. Our previous studies have shown that the optical bandgap of Sn-doped 100-nm polycrystalline -Ga 2 O 3 was ≈4.65 eV, [49] and detected the light in UV C-band (UVC, 100-280 nm). Therefore, a 265-nm light-emitting diode (LED) was selected for the optical stimulation of synaptic transistors. The EPSC of a synaptic transistor induced by a single optical spike incident on an Sn-doped 100-nm -Ga 2 O 3 channel is shown in Figure 2a. The light pulse was sustained for 0.5 s with a power density of 2.84 μW cm −2 . Following the optical pulses, the EPSC reached a maximum of 73 nA (EPSC max = 73 nA) at the end, maintained a slow decay state, gradually decayed ≈200 s, and returned to their initial values. Similarly, synaptic transistors can also be activated by electrical stimulation. Both IPSC and EPSC can be induced by applying positive and negative electrical spikes to the highly p-Si back-gate of the synaptic device, as shown in Figure 2b,c, respectively. Electrical stimulation was performed by applying 50 V (positive) and −30 V (negative) for 0.5 s. When an electrical spike with positive V G is applied, I D first increases and then decreases slightly. In contrast, when an electrical spike with negative V G is applied, I D first decreases and then increases slightly. This behavior is similar to the depolarization of membrane potentials in biological neuronal cells. [50] The recovery time is 7.48 s for positive V G and 8.99 s for negative V G . Short-term memory is stored in the hippocampus and can last from a few hundred milliseconds to a few minutes, depending on the concentration of the rememberer. [51] Considering the current recovery time in the electrical stimulation mode of the proposed device, the short-term memory formation process can be mimicked. Whether PSCs were excitatory or inhibitory was defined by the amplitude of PSCs just before the termination of the electrical pulse. The induction of IPSC and EPSC by positive and negative electric pulses is related to the charging and discharging of electrons to the Ga 2 O 3 /SiO 2 interface trap after the pulse is sustained. Figure 2e,f illustrate the energy band diagrams of the negative and positive surface states generated immediately after the positive and negative pulses, respectively.
When a positive V G is applied, the energy band of the Ga 2 O 3 thin film near the Ga 2 O 3 /SiO 2 interface is bent downward and the surface state of the gate oxide decreased. This causes electrons to accumulate near the Ga 2 O 3 /SiO 2 interface and the value of I D increases at the start of a positive electrical spike. The acceptor surface state of the gate oxide becomes negatively charged when it is downshifted below the Fermi level of the Ga 2 O 3 film. This weakens the positive V G effect on the band bending of the Ga 2 O 3 film and reduces the electron accumulation near the interface resulting in a subsequent decrease in I D during the electrical spike. At the end of the positive pulse duration, the interface temporarily remains in an acceptor-like state with a negative charge, which has an instantaneous effect after returning to zero gate bias. The negative interfacial state slightly depletes electrons near the Ga 2 O 3 /SiO 2 interface. Subsequently, the synaptic current decreases compared to the initial value (initial IPSC) when V G = 0 V. As the negatively charged acceptor-surface state gradually re-turns to its initial state, the synaptic current of the device returns to its initial level.
When a negative V G is applied, the energy band of the Ga 2 O 3 thin film near the Ga 2 O 3 /SiO 2 interface bends upward and the surface state of the gate oxide shifts upward. This contributes to the initial decrease in I D when electrons near the Ga 2 O 3 /SiO 2 interface are depleted and a negative electric spike begins. The donor-surface state of the gate oxide becomes positively charged when it rises above the Fermi level of the Ga 2 O 3 thin film. This counteracts the effect of negative V G on electron depletion near the Ga 2 O 3 /SiO 2 interface. Consequently, I D increases during the electrical spike. At the end of the negative electrical spike, the transiently remaining positively charged donor-surface state slightly accumulates electrons near the Ga 2 O 3 /SiO 2 interface, resulting in the EPSC of the Ga 2 O 3 synaptic transistor. As the positively charged donor-surface state gradually restores to its initial state, the EPSC of the device returns to its initial current level. The measured voltages are +50 V for positive V G and −30 V for negative V G and they are related to the Fermi level alignment of each layer of the material stack in this device. In the process of Ga 2 O 3 and p + -Si being aligned so that the Fermi level is even, Ga 2 O 3 is essentially in equilibrium with the energy band at the Ga 2 O 3 /SiO 2 interface being upwardly bent, which is why a relatively larger voltage magnitude is required for a positive gate spike that bends the band down. The behavior of the carriers in the channel when such spikes are applied can be explained using the schematic energy band diagram models ( Figure S3, Supporting Information) for each step. Figure 2g illustrates the results of measuring and collecting the postsynaptic current corresponding to the pulse duration of optical stimulation (L), positive electrical stimulation (E + ), and negative electrical stimulation (E − ); time-dependent raw data are presented in Figure S4, Supporting Information. L and E − , which induced synaptic strengthening, exhibit positive values. In the case of L, the postsynaptic current increased linearly as the pulse duration increased, whereas E − demonstrates a tendency to saturate following an initial increase. Furthermore, because E + is a synaptic inhibitory stimulus, it has a negative value, and as the duration increases, the postsynaptic current decreases and finally saturates.
The PPF characteristics of synaptic transistors stimulated by two successive optical and electrical pulses are examined ( Figure 3a-c)): the pulse duration and interval (Δt) are 0.5 s. When successive light pulses were applied, it was observed that the second light pulse induced a larger EPSC than the first (EPSC1 < EPSC2), and PSC induction by positive and negative electrical stimulation also exhibited similar trends. Figure 3d illustrates the relationship of the PPF index ((PSC2/PSC1) × 100) to the Δt of the stimulus, obtained from the time-dependent synaptic currents ( Figure S5, Supporting Information). The PPF index was fitted using the equation PPF = 1 + C 1 exp(-t/ 1 ) + C 2 exp(-t/ 2 ) to obtain the 1 and 2 decay times associated with the reduction of the PPF index of the optical and electrical pulses. [52,53] The associated 1 and 2 decay times for the optical pulse, positive electrical pulse, and negative electrical pulse are 0.09 and 2.42, 2.69 and 2.69, and 8.37 and 1.01 s, respectively. Notably, these decay time categories are consistent with the time scale of biological synapses, thereby demonstrating that the proposed synaptic transistor can be used for neuromorphic computing and operation. [25,54] The energy consumption of artificial synaptic devices is one of the most fundamentally important characteristics, and the consumption of each device per synaptic event can be calculated using the formula E = P · I · t d (where P denotes the pulse amplitude, I indicates the triggered postsynaptic current, and t d denotes the pulse duration). Generally, the energy expenditure of biological synapses is ≈1-10 fJ per synaptic event. [6,55] The minimum amount of energy consumed during stimulation of the proposed Ga 2 O 3 -based synaptic transistor is 0.81 nJ, which is rather high in this study. Energy consumption depends on the pulse duration, driving voltage, and device size. As demonstrated in previous studies, [49] the driving voltage can be reduced by optimizing V TH via depletion region control. Additionally, energy consumption can be reduced by optimizing pulse measurement and the maturation of Ga 2 O 3 device process technology, such as improving subthreshold swing (SS) by reducing the contact resistance of the transistor and device scale-down.
Subsequently, the long-term plasticity of synapses was investigated by observing changes in the synaptic weight during repetitive stimulation. Long-term potentiation (LTP) and long-term depression (LTD) are associated with long-term plasticity with either an increase or decrease in synaptic weight, respectively. [56] The following results demonstrate that the long-term plasticity of synaptic -Ga 2 O 3 phototransistors can be realized. Figure 3e presents the absolute value of the postsynaptic current change (|ΔW|) corresponding to the number of consecutive spikes. The function ΔW is calculated using the expression ΔW = ΔPSC/PSC0, where PSC0 denotes the initial synaptic current and ΔPSC denotes the current of the second spike minus the current immediately at the end of the first spike. Moreover, |ΔW| increased as the number of spikes accumulated when optical and electrical stimulations were applied. Increasing the spike number, |ΔW| can be tuned from ≈260% to 1730% and from ≈57% to 183% for optical and electrical stimulation, respectively. Generally, LTP can be transitioned through repeated stimulus rehearsal events from short-term plasticity (STP). [57] As shown in Figure  S6, Supporting Information, when 50 spikes versus single spike stimulation were applied, the current continued to be enhanced or suppressed 100 s after stimulation, confirming a slow recovery. In other words, this implies a delay in recovery time, suggesting a switch from STP to LTP. The |ΔW| corresponding to the spike frequency is presented in Figure 3f: the detailed extraction is provided in Figure S5, Supporting Information. In each measured frequency range, |ΔW| increases from ≈120% to 200% and from ≈10% to 49% for optical and electrical stimulation, respectively. These results validate the LTP of synaptic -Ga 2 O 3 phototransistors under light and -V G spike stimulation and the LTD of the device under +V G spike stimulation. Consequently, it can be interpreted that the influence of optical stimulation for LTP in the proposed device is more significant than that of electrical stimulation for both LTD and LTP.
Spike-timing-dependent plasticity (STDP) is widely considered a key function of the biological nervous system. [58] STDP exhibits synaptic weighting properties that can be modulated by the relative timing between pre-and postsynaptic spikes (Δt pre − post ). Generally, the STDP of biological synapses can be classified into two types: symmetric and asymmetric Hebbian learning rules, which are expressed as. [59,60]  where A and denote the scaling and time constants, respectively, and ΔW represents the relative change in synaptic weights.

ΔW = Aexp
Two correlated spikes with different waveforms and interval times were applied to the device, and Figure 4 illustrates the fitting results of the synaptic-weight change versus Δt for four measurements. The coexistence of LTP and LTD in synaptic -Ga 2 O 3 phototransistors enables different types of STDPs through a combination of different stimuli. Figure 4a-f shows the STDP of a synaptic -Ga 2 O 3 phototransistor under combined photostimulation (L), electrical stimulation with positive V G (E + ), and electrical stimulation with negative V G (E − ). The time-dependent raw data for these measurements are provided in Figures S5 and  S7, Supporting Information. It has been found that four types of Hebbian learning rules can be realized in synaptic -Ga 2 O 3 phototransistors. [61] Moreover, when both presynaptic and postsynaptic spikes are either L or E − , ΔW is always positive, which indicates symmetric distributions for Δ t pre − post = 0 (Figure 4a,c). This represents the symmetric Hebbian learning rules for the device. When both pre-and postsynaptic spikes are E + , ΔW is always negative with symmetric distributions for Δ t pre − post = 0 (Figure 4b), indicating a symmetric anti-Hebbian learning rule for the device. For the combination of E + pre-and E − postsynaptic spikes, ΔW is positive for Δt pre − post > 0 and negative for Δt pre − post < 0 (Figure 4d). For the combination of L pre-and E + postsynaptic spikes, ΔW is negative for Δt pre − post > 0 and positive for Δt pre − post < 0 (Figure 4e). As the absolute value of Δt pre − post decreases, potentiation (ΔW > 0) or depression (ΔW < 0) improves; thereby enabling asymmetric Hebbian learning. Finally, as shown in Figure 4e,f, the asymmetric Hebbian rule was only confirmed in the excitatory (L) and inhibitory (E + ) configurations in the optical and electrical stimulation combinations. The result is caused by the fact that both stimuli (L and E − ) are reinforcers; however, the ranges of weight modulation are different. The Hebbian rule can be achieved by optimizing the light intensity or voltage amplitude. The STDP time window ( ) of the synaptic -Ga 2 O 3 phototransistor was calculated by fitting the experimental data to the equation above, [56] and it was found that varied from 0.02 to 3.29 s depending on the configuration of the pre-and postsynaptic spikes. The values obtained from the fitting are listed in Table  S1, Supporting Information. Interestingly, the difference in the decay time of the postsynaptic current between optical and electrical stimulation was significant (Figure 2a-c). Therefore, when both optical and electrical stimulations are used to implement STDP, the values are different for Δt pre − post > 0 and Δt pre − post < 0 (Figure 4e). The weight-enriched STDP characteristics lead to various learning rules. The adoption of different learning rules in artificial neural networks enables those with improved learning efficiency to handle more complex scenarios. [61] Importantly, we note that the coexistence of optical and electrical stimulation in synaptic -Ga 2 O 3 phototransistors enables homeostatic feedback regulation, facilitating synaptic plasticity realization. These results indicate that ΔW can be programmed using synaptic -Ga 2 O 3 phototransistors if the stimuli are properly matched.
Subsequently, inspired by the visual information processing mechanism of a bee, we designed the proposed synaptic -Ga 2 O 3 phototransistor to mimic the typical neurological behavior, the associative learning process. Bees can perceive UV light and use it to form visual memories. Figure 5a illustrates the process in which visual information from a bee's retina is supplied to the optic lobe via the optic ganglion as a pathway for neurons. [62] The associative learning process was emulated using optical and negative electrical stimulation, a combination of excitatory stimulation types. Associative learning refers to the induction of an unconditioned response (UR) via a conditioning process in which a neutral stimulus (NS) that elicits no response is associated with an unconditioned stimulus (US) that elicits a UR. The bee does not respond to UV light (NS), and navigating flowers (UR) is induced only during feeding (US). At the time, through the conditioning process, the bee can navigate flowers even when only UV light is visible. In the example, optical stimulation was emulated with NS and negative electrical stimulation with US. In this measurement, the intensity of the light source was adjusted to 0.31 μW cm −2 to ensure that the weight modulation range of the optical stimulation is the same as that of the electrical stimulation. We arbitrarily set the threshold at which the UR is generated to be ≈14 nA. When only optical stimulation was applied, EPSC responses were below the threshold, indicating "no response" (Figure 5c). In contrast, when only electrical stimulation was applied, the current exceeded the threshold and "navigating flowers" occurred ( Figure 5d). Following that, the conditioning process was performed by associating the UV light with feeding by applying optical and electrical stimulations, as shown in Figure 5e. Following this task, optical stimulation could independently elicit above-threshold EPSC responses, indicating navigating flowers (Figure 5f). Comparing Figure 5c and Figure 5f, it is evident that associative learning works effectively because a larger EPSC is induced after the conditioning process, even under the same stimulation.
In addition, the fabricated synaptic -Ga 2 O 3 phototransistors can be used as the components (artificial synapses) of a deep neural network (DNN). In a DNN, the currents of artificial synapses, that is, the synaptic weights, are either reinforced or degraded to appropriate values using a back-propagation algorithm. [63,64] Therefore, artificial synapses should exhibit linear and symmetric weight update characteristics to demonstrate high learning accuracy. [63,65] To investigate the potentiation/depression characteristics of the fabricated synaptic -Ga 2 O 3 phototransistor, 18 optical (L) and electrical stimulation pulses (E + ) with the same duration and interval were applied (Figure 6a).
At the end of the stimulation, the drain current was recorded as a function of time and potentiated/depressed by L/E + stimulation. The conductance extracted from the current after the end of each pulse is plotted as a function of the pulse number, as shown in Figure 6b: the conductance was calculated from the drain current at V D = 10 V. To evaluate the applicability of the proposed device as an artificial synapse, the linearity was extracted using the calculated conductance. The linearity of potentiation and depression ( p , d ) were extracted by fitting according to the equations provided by Jerry et al. [66] and the results were p = 0.10 and d = -3.89. This suggests that the synaptic -Ga 2 O 3 devices may have weight-modulating characteristics that can be realized using ANNs. It is generally known that ANNs with high computational accuracy can be achieved with small nonlinearities (< 0.5 to 1). [65] The proposed synaptic -Ga 2 O 3 device exhibits good linearity in the potentiation region while exhibiting relatively less linearity in the depression region. It is expected that this can be improved in future research by optimizing the pulse composition and adjusting the duration. Based on the weight-modulating properties of the -Ga 2 O 3 device, we performed MNIST recognition training using a CNN model to demonstrate the applicability of synaptic -Ga 2 O 3 phototransistors in neuromorphic computation. The developed CNN model includes two convolutional layers and two dense layers, as depicted in Figure 6c. Through weight update in each layer, the corresponding CNN model learns the MNIST data set. Once learning is completed, the weight values of the two dense layers are quantized. We used the conductance values of the proposed synaptic -Ga 2 O 3 devices extracted from 18 optical and electrical stimulation pulses for weighted quantization of the dense layer. As shown in Figure S8, Supporting Information, 18 conductance values are normalized to a value between −1 and 1, and the weight of the dense layer is quantized with respect to the normalized conductance value. The weights of the quantized dense layer were inserted into the CNN model again, and the pattern recognition accuracy was tested. The test results exhibited an accuracy of 96.92%.

Conclusion
In this study, hybrid optoelectronic synaptic transistors based on Sn-doped polycrystalline -Ga 2 O 3 were fabricated. The proposed device was stimulated using optical spikes in the UV region and gate-voltage spikes and mimicked several crucial synaptic functions of EPSC, IPSC, PPF, SNDP, SRDP, and STDP. It has been experimentally confirmed that optical stimulation is closely related to the photogate of synaptic -Ga 2 O 3 phototransistors and electrical stimulation is caused by the gate-dielectric interface state affecting the behavior of carriers. Notably, the synergistic effect of optical and electrical stimulation implemented in this device is applied to the visual memory formation process of bees to successfully mimic classical associative learning. Furthermore, image recognition of handwritten images was performed using the synaptic -Ga 2 O 3 phototransistors. Using these optical and electrical interconnections as neurons, all-optical and electrical stimulation synaptic devices can be connected to form an integrated optoelectronic neural network. Moreover, the findings of the present study encourage the development of highperformance optoelectronic synaptic devices for neuromorphic computing by exploring a rich library of semiconductor materials because -Ga 2 O 3 is a type of oxide semiconductor thin film emulated in UVC. Interestingly, this study presents an avenue for the application of oxide-based optoelectronic devices in photonics and neuromorphics. Furthermore, the successful implementation of the fundamental functions of these biological synapses suggests the utility of Ga 2 O 3 -based optoelectronic devices for next-generation neuromorphic computing. Future studies can aim to develop a device capable of accumulating and reducing holes and electrons in a channel by using the difference in polarization or wavelength of incident light to operate on any signal combination of optical and electrical spikes.

Experimental Section
In this study, gallium oxide thin films (300 nm) were deposited on p + -Si/SiO 2 wafer substrates via radio frequency (RF) sputtering using a 99.99% pure Ga 2 O 3 target. The deposition process was performed in an inert atmosphere of Ar gas at a pressure of 1.3 mTorr and 70 W of power. The deposited thin film of an amorphous phase was crystallized into monoclinic Ga 2 O 3 ( -phase) by annealing at 900°C for 1 h in a tube furnace under ambient atmosphere; the furnace temperature was increased from ambient temperature at a ramp rate of 10°C min −1 , held at 900°C for 1 h and then cooled again. The intrinsic -Ga 2 O 3 thin film transformed into an n-type semiconductor through spin-on-glass (SOG) Sn doping technology. [67] In the SOG process, the Sn dopant source solution was prepared by stirring a mixed solution of 0.0975 g SnCl 4 (99.995%, 2.226 g mL −1 ) in 20-ml 2-methoxyethanol (CH 3 OCH 2 CH 2 OH, 2MOE) at 60°C for 1 h. The solution was spin-coated on the -Ga 2 O 3 surface at 3000 rpm for 60 s and subsequently heated on a hotplate at 120 and 300°C for 10 min. A two-step baking process was adopted to prevent cracking during solvent removal of the SOG layer. Subsequently, Sn atoms contained in the SOG film were diffused into the polycrystalline -Ga 2 O 3 thin film by annealing (450°C in N 2 ) for 1 h. It was observed that Sn doping concentration was ≈1 at.% by energy dispersive X-ray spectroscopy (EDX) mapping. [67] The crystal structures of the Sn-doped polycrystalline -Ga 2 O 3 thin films were investigated using a transmission electron microscope (TEM, EM-ARM200F, JEOL, Ltd., Japan) at 300 keV. Fast Fourier transform (FFT) patterns were analyzed using Gatan DigitalMicrograph® (AMETEK, Inc., USA) for lattice parameter acquisition and compared with data from monoclinic gallium oxide JCPDS (card number 01-076-0573). The crystal structures of Sn-doped -Ga 2 O 3 thin films fabricated by SOG doping were characterized by TEM analysis. Electron diffraction patterns were obtained from the TEM crosssectional image analysis data ( Figure S9, Supporting Information). The cross-sectional image confirmed that the Sn-doped -Ga 2 O 3 thin film was 100-nm thick. The FFT pattern analysis of the Sn-doped 100-nm -Ga 2 O 3 thin film revealed several planes. Their lattice parameters were compared with JCPDS data (Table S2, Supporting Information), and the lattice constants of four planes, viz., (311), (−311), (−401), and (110), were observed to be similar to those of JCPDS. This can be interpreted to imply that no detectable mechanical strain was induced in the Sn-doped 100-nm -Ga 2 O 3 layer. However, the lattice parameters of the Sn-doped 100-nm -Ga 2 O 3 exhibited a consistent upward trend with a doping-dependent increase of 0-0.3%, which is identical to the Sn-doping-induced lattice change observed in previous studies. [68] Sn was doped by substitution at the Ga site, and because the atomic radius of Sn is larger than that of Ga, a slight lattice expansion was induced. Accordingly, the lattice parameters may be the same or increased in comparison to the intrinsic parameters.
The Sn-doped polycrystalline -Ga 2 O 3 synaptic FET was fabricated with a back-gate structure by forming source/drain (S/D) electrodes of Ti/TiN (5/200 nm) via RF sputtering on top of the SOG-doped sample. The patterning of the electrodes was performed using conventional photolithography followed by a lift-off process. The p + -Si and SiO 2 were used as the back-gate electrode and gate dielectric, respectively. Finally, the fabricated back-gate FET was annealed in air at 400°C for 1 h to improve the contact resistance. The electrical and photocurrent characteristics of Sn-doped polycrystalline -Ga 2 O 3 synaptic FETs were investigated using a semiconductor parameter analyzer (Keithley 4200-SCS, Tektronix, USA) under UVC irradiation. A commercially available UV LED source (Mounted LED, M265L3, Thorlabs, USA) with operating wavelengths in the range of 200 nm (38 μW cm −2 ) to 300 nm (38.6 μW cm −2 ) with a peak emission wavelength of 265 nm (51 μW cm −2 ) was used as the UVC source for this work. The time-dependent optical signal was controlled as an input transistor-transistor logic (TTL) signal using a function generator (AFG-2225, GW Instek, Taiwan) connected to the UVC LED source.

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