Photonic Synapses for Image Recognition and High Density Integration of Simplified Artificial Neural Networks

With the rapid development of artificial intelligence (AI), there is an urgent need for developing a biological sensory perception system that can simulate the human brain for information processing. Inspired by the biological vision system, photo‐responsive photonic synapses are ideal devices for constructing photosensitive artificial neural networks for neuromorphic computing tasks. This paper reports a stable photonic synaptic device in an array layout with adjustable synaptic plasticity under ultraviolet light pulses. Since the heterojunction has a photoconductivity effect and the trap layer provides superior charge carrier trapping capability, optical sensing, memory, and neuromorphic computing are integrated into a single device. Meanwhile, supervised learning of handwritten digitals is achieved by exploiting the multistate conductance by photoelectric co‐modulation and the specific decay law. The recognition rate reaches 90.6% and hardly changes with time. Additionally, the device can simplify the artificial neural network (ANN) and reduce its size to 3.78% of the original network while retaining strong fault tolerance and learning ability. The photonic artificial synapses based on ultraviolet light modulation provide a novel and effective approach for photosensory ANNs to perform in situ computation.

strength in practical production. [32,33] Both the NiO film and ZnO film are important optical materials, and they are frequently applied in optical devices, photodetectors, and lightemitting diodes [34,35] owing to their good absorption ability of ultraviolet light. Reducing the composite center of materials or adding the trap layer to capture charge carriers helps to improve the memory capacity of synapses, i.e., prolonging the lifetime of photocarriers in optical materials. [36,37] Besides, it is worth noting that using heterojunction as dielectric layers will combine the properties of the two materials, thus further optimizing the characteristics of the photon synapses and providing new possibilities for the integrated system of photosensitive storage and computing.
Herein, we report a stable photonic artificial synaptic device based on NiO/ZnO heterojunction, which presents the array layout with the structure of Si/SiO 2 /Al/ZnO/NiO/indium tin oxide (ITO). Through the modulation of light pulses, the device can imitate the important behaviors of biological synapses such as short-term plasticity (STP), long-term potentiation (LTP), paired-pulse facilitation (PPF), and forgetting rule, attributed to the photoconductivity effect of heterojunction and the excellent ability of AlO y trap layer to capture holes. The synaptic plasticity of photoelectric co-modulation was employed to simulate a three-layer ANN. The ANN can supervise the learning of handwritten digits in the Mixed National Institute of Standards and Technology (MNIST) data set. [38] The recognition rate reached 90.6% and remained stable with time. Additionally, inspired by reservoir computing, [39] the network scale of the three-layer ANN was considerably reduced by adopting the dynamic evolution and mapping approach for letter images. [40] While ensuring strong fault tolerance and learning ability, the recognition rate can reach 95%. The construction of photonic artificial synapses based on ultraviolet light modulation provide a new approach to the high-density integration of neuromorphic systems. Figure 1a shows the schematics of the human visual system and the photonic artificial synapse. The optical information is received from the environment by the retina, sent to the visual cortex in the human brain, and then transmitted by synaptic structures between the neurons. In this synaptic structure, Si/ SiO 2 was used as the substrate, and an ultraviolet lamp represented the front of the synapse, while the top electrode ITO corresponded to the receptor at the front of the synapse, and the current between the top electrode and the bottom electrode Al was the post-synaptic current (PSC). The dielectric layer ZnO was grown by atomic layer deposition (ALD), and the dielectric layer NiO was deposited by physical vapor deposition (PVD). More detailed information on device fabrication will be presented in Experimental Section.

The Photonic Artificial Synaptic Device
The synaptic unit shows an array layout with a large area, and atomic force microscopy (AFM) and X-ray diffraction (XRD) images reveal that both the ZnO film and NiO film have higher quality after high-temperature treatment (Figure 1b,e, and Figure S1, Supporting Information). To demonstrate the thickness and elements distribution of each layer in the device, transmission electron microscopy (TEM) and energy dispersive X-ray (EDX) were employed to obtain cross-section images and EDX mapping results with the line-scan pattern (Figure 1c,d). Besides, the Al electrode was oxidized to AlO y due to the influence of oxygen.
The typical current and voltage (I-V) curves under dark and light conditions are illustrated in Figure 1f. Positive voltage sweeps (0→500 mV→0) and negative voltage sweeps (0→-500 mV→0) were applied to the top electrode, and the current of the device under light conditions increased significantly. Meanwhile, the conductance of the device gradually dropped after ten consecutive positive and negative voltage sweeps at the same magnitude, indicating a weakened synaptic plasticity behavior (see Figure S2, Supporting Information). Then, the voltage range was increased to ±2 V, and the current was measured under dark and light conditions respectively. Obvious memristive behaviors can be observed between 1 and 2 V, realizing the storage of photogenerated carriers (see Figure S3, Supporting Information). Figure 1g presents the reliability of the device under dark and constant light over 1000 s.

Tunable Synaptic Plasticity
The synaptic weight refers to the strength of connections between synapses. It may increase or decrease as the brain receives new information, which is a phenomenon known as synaptic plasticity. [41] In this study, artificial synapses based on NiO/ ZnO heterojunctions were triggered by light pulses. Figure 2a illustrates the PSC under eight different durations with a constant light intensity of 18.3 µW cm -2 . The PSC increased with the light pulse width. In Figure 2b, the light pulse width was kept at 1 s, and the light pulse intensity was adjusted. The results indicated that as the intensity of the light pulse increased, ∆PSC (i.e., the change of the PSC of the device) gradually increased. Meanwhile, the device responded to the light pulse with a wavelength range of 260-380 nm (the step size is 10) (see Figure S4, Supporting Information). The variations of the widths, intensities, and wavelengths of the light pulse can alter the photocarrier concentration of the device, thus adjusting the synapse weight. As shown in Figure 2c, the device can be transformed from STP to LTP by changing the number of light pulses within 100 s. When five pulses of 1s were applied (the pulse frequency was 0.05 Hz), the PSC approached the initial current of 150 pA after 100 s. When the number of pulses was raised to 10 (the pulse frequency was 0.1 Hz), the PSC did not return to the initial current value after 100 s, and the conductance was still higher than 1.75 nS after 550 s (see Figure S5a, Supporting Information). After 100 s, the photogenerated carriers were still not entirely recombined when only one light pulse was provided (see Figure S5b, Supporting Information). By employing a negative voltage pulse of 100 mV with a pulse width of 10 ms and frequency of 1 Hz, the conductance could decline to the original value (see Figure S5c, Supporting Information). It can be observed that by using the negative voltage pulse, the photogenerated carrier recombination process can be expedited, and the synaptic weight was reduced. The experimental principle will be further discussed below. The synapse weight can be modified freely through www.advelectronicmat.de coordinated control of light and voltage pulses, which lays the foundation for simulating the synapse functions in the human brain.
PPF is a significant short-term phenomenon, [42] which indicates that when two excitatory presynaptic pulses are applied in succession, the second pulse generates a larger excitatory post-synaptic current (EPSC) than the first one. Meanwhile, the EPSC induced by the second pulse is affected by the time interval (∆t) between the two pulses, and a larger time gap will result in a smaller magnitude of increase in EPSC. In Figure 2d, A1 and A2 are the currents induced by the two continuous optical pulses, ∆t is the time interval, the optical pulse width is 1 s, and the intensity is 38.2 µW cm -2 . As shown in Figure 2e, the experimental data fit the curve with the double exponential decay function [43] : where τ is the characteristic relaxation time, and C 0 is the degree of initial facilitation. As ∆t rose from 100 ms to 1000 ms, the PPF ratio (defined as A2/A1 × 100%) decreased from 155 to 117%. When the time interval between the two light pulses was further increased to 100 s, A2/A1 is still higher than 100% (see Figure S6, Supporting Information). The learning and forgetting process of the human brain can be imitated by exploiting the long-term memristive capability of the device. It can be seen from Figure 2f that the first learning process consisted of 20 light pulses, followed by a spontaneous drop of 100 s (without light). Then, eight pulses of light were applied as the second learning process, and PSC dropped naturally again.

www.advelectronicmat.de
The PSC of the second forgetting curve was higher than that of the first one when the spontaneous decline time was the same, which conforms to the rule of the Ebbinghaus forgetting curve. [44] Figure 2g demonstrates the change of the band diagram and the movement process of carriers of the device under and after the removal of illumination. When ZnO was grown by ALD, aluminum oxides were generated due to the oxidation of the Al electrode surface to a certain extent in an oxygen environment. Meanwhile, Al was also oxidized through direct contact with ZnO. However, Al was not totally oxidized, but produced AlO y , [36,37] which formed a large number of defect states, and ZnO adjacent to the bottom electrode turned into ZnO x . Both the NiO/ZnO interface and the ZnO x /AlO y interface had built-in electric fields in the equilibrium condition. The defect states in ZnO were greatly reduced by the high-temperature annealing method after the deposition of NiO. In this case, the weak n-type ZnO was presented, so the NiO/ZnO interface had www.advelectronicmat.de a weak built-in electric field. When exposed to light, the electrons in the valence band of NiO and ZnO transitioned to the conduction band, leaving holes in the valence band, and some electrons were captured by AlO y traps. With the increase of photocarriers, PSC steadily increased toward saturation. Electrons gradually transitioned back to the valence band when the light was turned off, and the captured holes in AlO y moved to ZnO x by overcoming the barrier due to thermal excitation, thus lowering the PSC. Then, the negative voltage pulse of 100 mV was applied. In this case, the speed of the hole transferred to ZnO x increased, the electrons were also trapped by AlO y , and the PSC dropped quickly because only a few carriers participated in the conduction. After the electric field was removed, the original initial current can be still maintained. NiO and ZnO created photocarriers due to the photoconductivity effect, and NiO strengthened the potential barrier for carrier transition in ZnO. In this way, NiO prolonged the relaxation time of photocarriers, and the device exhibited photoelectric doubletunable multistate conductance. When the ZnO layer was removed or the bottom electrode Al was replaced, the optical properties of the device were changed, which was not conductive to the simulation of photonic synapses (see Figure S7, Supporting Information).

Image Recognition with Strong Fault Tolerance
The ANN model mimics the biological neural network (especially the brain), and it consists of a large number of connected artificial neurons to perform calculations. The traditional neural network based on synaptic devices is controlled by electrical impulses, and the tunable conductance of electronic synapses can be exploited to simulate the synaptic weight, thereby realizing the function of neuromorphic computing. However, the multistate conductance is not stable after electrical modulation, which influences the identification accuracy. In this study, by considering the characteristics of the photonic artificial synapse, a dynamic recognition approach was proposed to effectively avoid the disturbance problem.
First, the regulation of optical stimulus on the weight of synaptic devices was explored. To ensure that the device possessed multistate conductance, 30 light pulses were given continuously at a constant duty cycle of 50%. As shown in Figure 3a, the device had 30 continually increasing conductance values at the pulse width of 250 ms. When the light was removed, the conductance of the device decreased naturally. When the pulse width was 500 ms, the device easily obtained saturation current (see Figure S8, Supporting Information), which makes it not suitable for application in neural networks. Furthermore, the conductance attenuation rule of the device under different pulse widths and numbers were investigated. As shown in Figure 3b, the optical pulse period was set to 500 ms, the pulse was applied continuously five times with a width of 100, 200, 300, and 400 ms, respectively. The attenuation curve was fitted by the Stretch Exponential Function (SEF) [45] (Figure 3c): where τ is the characteristic relaxation time reflecting the decay process. Figure 3d demonstrates that τ was almost unaffected by the width of the light pulse and remained stable at ≈12 s. However, τ rose significantly with the number of light pulses (the pulse width was 500 ms) (Figure 3e,f). Figure S9, Supporting Information, shows the specific change of ∆PSC. Therefore, we constructed the dynamic identification method according to the optical characteristics of the device. The 30 conductance values ( Figure 3a) were divided into three groups (each group had 10 conductance values) with current decay curves of three τ values in terms of the pulse number, and the initial conductance value remained constant. The attenuation curves of pulse number of 5, 15, and 25 were used in the three groups ( Figure 3e). In this way, the weight of synapses and their relationship with time after modulation by light pulse were obtained. Figure 3g,h present the diagram of a three-layer ANN for image recognition. The 256 input neurons correspond to pixels of an image taken from the MNIST handwriting dataset, and the 10 output neurons corresponded to 10 numbers from 0 to 9. To verify the fault tolerance of the network, 0-70% salt-and-pepper noise was randomly added to the digital image. The number "5" (28 × 28 pixels) was used as an example (see Figure S10, Supporting Information, shows the complete noise image). Then, the bilinear interpolation function was adopted to obtain the input image of 16 × 16 pixels. The number of hidden layers was determined to be 64. The relationship between recognition rate and training epoch without noise was obtained by varying the number of hidden layers (see Figure S11, Supporting Information). 1000 training images and 50 test images were used for each number. Figure 3i illustrates the flow chart of the back-propagation (BP) algorithm used to build the network in one epoch (X, Y, and Z represent the input layer, hidden layer, and output layer, respectively). In one epoch, there are 10000 training images. For each input of the training image, the weight in the network is updated according to the calculated error rate. Then, the recognition rate is figured out by using the test image after training. Generally, it takes many training epochs to obtain a relatively high recognition rate, and the initialization of V and W only happens before the first epoch.
The change curves of the recognition rate during 1000 training epochs are presented in Figure 4a. After 1000 training epochs, the recognition rate still reached 77% with 40% noise (Figure 4b). Subsequently, to test the suitability of the photonic artificial synapse in ANN, the weights in the neural network were replaced with 31 conductance values obtained by optical pulse modulation. The results indicated that the recognition rate of the device was 90.6% after 1000 training epochs, showing its good image recognition capability (Figure 4c). As shown in Figure 4d, the recognition rate of the photonic artificial synapse did not change considerably with time, showing a strong fault tolerance. Figure 4e illustrates the output value of the digit "0" in the training process of 1000 epochs. With the increase of the training epoch, the output value of number "0" was much higher than that of other numbers, and the other nine numbers also exhibited a similar trend (see Figure S12, Supporting Information). It can be seen from the mean confusion matrix that during the learning process of the neural network, the color of the diagonal gradually deepened, and all numbers were www.advelectronicmat.de identified (see Figure S13, Supporting Information). According to the conductance distribution in the interval of 1.3-2.1 ns in Figure 4f, the conductance became increasingly dispersed with the increase of the training epoch. Figure S14, Supporting Information, intuitively shows the fitting results based on the Gaussian distribution curve. The synaptic weight can be modulated by the negative voltage pulse to return to the initial state after recognition. In conclusion, the photonic artificial synapse

www.advelectronicmat.de
is potential to be applied to ANNs owing to its excellent image recognition ability.

Construction of a Simplified Artificial Neural Network
When images are recognized by a three-layer ANN, the number of input layers is usually equal to the total pixels of the image, which makes the scale of the neural network difficult to reduce. A simplified ANN method was proposed by considering the devices' photoconductance characteristics while guaranteeing the recognition accuracy, as shown in Figure 5a. An image of 32 × 32 pixels (1024 pixels in total) was mapped into 32 pixels to represent the information of each line after dynamic evolution. The image recognition process was accomplished after the neural network transmission of the input layer (32 layers), hidden layer (24 layers), and output layer (7 layers). When the number of hidden layers was the same, the simplified network area was 3.78% of the original network area (set the number of hidden layer as x, then the ANN area ratio was 39x/1031x = 3.78%), which makes it possible to further improve the integration density of the chip. Additionally, ANN's computing speed was also improved.
Seven types of English capital letters images were selected from the dataset, [40] and each letter included 100 images with different fonts (such as italic, bold, deformed, etc.), among which 80% were employed for training and 20% for testing. Part of the images are shown in Figure S15, Supporting Information. Based on the gray level of the image, the pixel points were divided into five grades from 0 to 4, where "0" represents pure white and "4" represents pure black. The characteristics of optical pulses represented by different levels were shown on the right of the image. Each cycle was 500 ms, and each row had 32 pixels, so the evolution time was 16 s. The variation of the encoding current after each pixel of the row in the example was demonstrated in Figure 5b. Each point only represented the current at the end of one cycle, not the actual current curve. The information of the image was preserved due to the excellent synaptic plasticity of the device under the optical stimulation. After normalization, 32 points representing 32 rows were input into the ANN. The ∆PSC and the attenuation curves for different grades were determined according to the real optical response of the device (Figure 3b), where the ∆PSC was set to the average current for 5 pulse cycles and the decay curve was the same (τ: 12 s).
The current did not increase any further when it reached saturation, so a saturated ∆PSC was set for the device. As shown in Figure 5c, when ∆PSC was up to 200 pA, the recognition rate was 95% after 1000 cycles of training (The 50 recognition rates are shown in Figure S16, Supporting Information); when the saturation value was dropped to 120 pA, the recognition rate remained unchanged. Figure S17, Supporting Information, shows the cumulative percentage of input weight. It can be seen that appropriately reducing the saturation current did not

www.advelectronicmat.de
affect the recognition accuracy, so the method is universal. The following results were obtained when the saturated ∆PSC was 200 pA. The number of hidden layers influenced training speed, but there was no discernible difference in recognition rate after 1000 training epochs (see Figure S18, Supporting Information).
The total weight number of the neural network was controlled within 1000, and the network scale shrank significantly (see Figure S19, Supporting Information). In Figure 5d, the output of the letter "A" was higher than that of other letters as the training epoch increased, and the other six letters can also be www.advelectronicmat.de accurately identified (see Figure S20, Supporting Information). The average confusion matrix can show the learning process of neural network well. In the learning process, the color of the diagonal lines deepened, indicating that the recognition rate was gradually improved for each letter (Figure 5e). Besides, a certain percentage of salt-and-pepper noise was introduced to the images to evaluate the anti-jamming ability of this method. Taking the letter "D" (32 × 32 pixels) and its noise images as examples, the results are presented in Figure S21, Supporting Information. Even though the noise ratio was 50%, the recognition rate was higher than 70%, as shown in Figure 5f. Moreover, the analysis of the curve with a noise ratio of 10% indicated that when the weight of the input layer of the neural network changed in a certain range, the accuracy still reached 90%, showing its application potential for image recognition.

Conclusion
In this work, a photonic artificial synaptic device was successfully fabricated based on NiO/ZnO heterojunction, in which the AlO y layer served as the trap layer. The trap layer can slow down the recombination speed of photogenerated carriers and significantly enhance the memristive ability of ultraviolet light, thus laying the foundation for simulating the synaptic structure in the human brain. Meanwhile, a three-layer ANN was conducted by utilizing the multistate conductance obtained by electro-optical modulation and the stable PSC attenuation law. The recognition rate of handwritten digits in the MNIST dataset was 90.6%, and it could reach 77% even if 40% salt-andpepper noise was added. The recognition rate hardly changed with time, which provides a basis for reliable information processing and high fault tolerance of the neural network and proves that the device has good image recognition ability. Furthermore, through the dynamic evolution process, the alphabet image dataset was mapped, simplified to 32 pixels, and then input into the three-layer neural network. Under the premise of recognition accuracy of 95% and strong anti-interference, the scale of the simplified neural network was 3.78% of the original neural network. After simplification, the ANN's area drastically decreased. Our work indicated that the two-terminal synaptic device supports photoelectric co-modulation and has the potential for large-scale integration, which makes it a promising candidate for neuromorphic computing applications.

Experimental Section
Device preparation: The Si/SiO 2 (200 nm) substrate was ultrasonically cleaned in acetone, ethanol, and deionized water for 10 min successively, and then it was dried with nitrogen. The Ti (10 nm)/Al (100 nm) were deposited on the substrate by physical vapor deposition (PVD) as the adhesion layer and bottom electrode. Next, the ZnO film of ≈20 nm was grown by atomic layer deposition (ALD) at 200 °C. Subsequently, the 35 nm NiO film was produced by PVD dc sputtering at room temperature. The target material was Ni, and the power was 120 W. The oxygen and argon ratio was 10%. Afterward, the device was annealed in an oxygen atmosphere at 550 °C. Then, photolithography was conducted to define the shape of the top electrode as regularly arranged rectangles with a side length of 80 µm. Finally, 100 nm ITO was deposited by the PVD method as the top electrode, and the final device was obtained after soaking in a degumming solution.
Electrical/Optical characterization: Electrical tests were carried out on the semiconductor parameter analyzer (Agilent B1500A) with a semiconductor pulse generator unit (SPGU) under dark circumstances. Meanwhile, light pulses were produced via a xenon lamp system, and a small voltage bias of 0.1 V was applied to read the synaptic current. All the electrical and optical characteristics were performed in the ambient atmosphere at room temperature. X-ray diffraction patterns of the film were obtained by using a Bruker D8 Advance instrument with a 2θ range from 20 to 65° in a step of 0.01°. Atomic force microscopy (AFM) was applied in the tapping mode.

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