CsPbBr3/graphene nanowall artificial optoelectronic synapses for controllable perceptual learning

The rapid development of neuromorphic computing has stimulated extensive research interest in artificial synapses. Optoelectronic artificial synapses using laser beams as stimulus signals have the advantages of broadband, fast response, and low crosstalk. However, the optoelectronic synapses usually exhibit short memory duration due to the low lifetime of the photo-generated carriers. It greatly limits the mimicking of human perceptual learning, which is a common phenomenon in sensory interactions with the environment and practices of specific sensory tasks. Herein, a heterostructure optoelectronic synapse based on graphene nanowalls and CsPbBr3 quantum dots was fabricated. The graphene/CsPbBr3 heterojunction and the natural middle energy band in graphene nanowalls extend the carrier lifetime. Therefore, a long half-life period of photocurrent decay - 35.59 s has been achieved. Moreover, the long-term optoelectronic response can be controlled by the adjustment of numbers, powers, wavelengths, and frequencies of the laser pulses. Next, an artificial neural network consisting of a 28 × 28 synaptic array was established. It can be used to mimic a typical characteristic of human perceptual learning that the ability of sensory systems is enhanced through a learning experience. The learning behavior of image recognition can be tuned based on the photocurrent response control. The accuracy of image recognition keeps above 80% even under a low-frequency learning process. We also verify that less time is required to regain the lost sensory ability that has been previously learned. This approach paves the way toward high-performance intelligent devices with controllable learning of visual perception.


Introduction
In the era of data torrents, the exponential growth of electronic computing power has led to deviations from Moore's Law. During large-scale computing, the von Neumann architecture using a processing unit and a separate storage unit encounters a "storage wall" bottleneck, owing to the limitation of the transmission rate between the processing unit and the storage unit [1][2][3][4]. On the other hand, human sensory neurons conduct in-memory processing. For example, a phenomenon named perceptual learning demonstrates that the ability of sensory systems to respond to stimuli can be enhanced through a learning experience, and less time is required to regain the lost ability that has been previously learned. It is a basic mechanism for a human to process input signals with extremely high rates and low energy consumption [5]. Consequently, the elimination of the processing-storage transmission has great potential to solve the limitations of von Neumann's architecture.
Currently, neuromorphic computing that mimics in-memory processing functions of the sensory neurons has attracted intensive research interest [6,7]. The simulation of a synapse, which is a key component of the sensory neuron has become a breakthrough point in improving the calculation speed and reducing power consumption [8]. Conventional artificial synaptic devices are operated based on the principle of resistance changes under electrical stimulations. Transistors [9,10], phase-change materials [11], and resistive-change materials [12,13] have been used to realize the in-memory behaviors [14]. However, the electrical synapses have limited bandwidths due to considerable interconnection problems including delays and power losses in the process of signal transmission, and thus the neuromorphic computing speed is limited [15]. In contrast, investigations have proved that optical stimulations can considerably broaden the bandwidth and alleviate the interconnection problem [16]. The devices of artificial optoelectronic synapses provide a promising pathway for neuromorphic computing [17]. More importantly, the optoelectronic synapses can accurately simulate retinal neurons, enabling high-accuracy image detection and recognition [18]. Since 80 % of the information in the external world is transmitted to the human brain through visual perception, the artificial optoelectronic synapses exhibit a noncontact input method for the development of optical wireless communications and operations [19].
Artificial optoelectronic synapses can be divided into four types -heterojunction channel (HC)-typed synapses [20][21][22][23], oxide semiconductor-based synapses [14,24], floating gate-based synapses [25,26], and circuit-based synapses [27,28]. Among them, HC-typed synapses have introduced additional laser absorbers (such as perovskites [29][30][31], chalcogens [32][33][34], and organic semiconductors [35]) to form a heterojunction with high-mobility semiconductors in the channel region, showing highly sensitive photoconductance properties [36]. Moreover, the range of the photoresponse spectrum can be easily controlled by using the absorber with an appropriate bandgap. Perovskite is one of the most promising candidate materials for HC-type optoelectronics because of its high carrier mobility [37] and tunable bandgap [38]. However, perovskite-based HCtype synapses usually exhibit a very short memory time due to the rapid recombination of the captured charges at the heterogeneous interface. The challenge to achieve a significant long-term greatly response limits the simulation of typical in-memory processing behaviors such as perceptual learning [39].
Herein, to realize perceptual learning on neuromorphic computing devices, we report HC-type optoelectronic synapses based on graphene nanowalls (GNWs) and all-inorganic extremely CsPbBr 3 perovskite quantum dots (QDs). CsPbBr 3 QDs are used as laser absorbers layers and GNWs as semiconductor transmission layers. The built-in electric field of the CsPbBr 3 QDs/GNWs heterojunction can separate the photogenerated carriers, and the natural defects in GNWs can confine the transferred carriers, thus substantially extending the carrier lifetime. Moreover, the vertical nanosheet structure of GNWs can effectively increase the contact area between graphene and CsPbBr 3 QDs to improve the memory duration of the photocurrent. Under the stimulation of a 530 nm laser pulse, an extremely high value of the half-time period of photocurrent decay up to 35.59 s can be measured. Through the utilization of an artificial neuron network (ANN) consisting of the GNWs/CsPbBr 3 QDs synaptic array, the typical characteristics of human perceptual learning can be demonstrated.

GNWs/CsPbBr 3 QDs optoelectronic synapses
A biological nervous system constitutes lots of interconnected neurons to simultaneously receive information and achieve distributed in-memory processing [40]. Figure 1a presents a schematic showing a typical synaptic structure of the neural network, which consists of basic parts -cell body, axon, and dendrite. Input signals generate action Fig. 1 A biological synaptic structure for perceptual learning and GNWs/CsPbBr 3 QDs artificial optoelectronic synapse. a A biological nervous system with a typical synaptic structure. b Schematic of a synaptic structure. c Schematic of human perceptual learning, in which the sensory ability can be enhanced through a learning experience. d Schematic of a GNWs/CsPbBr 3 QDs artificial optoelectronic synapse. e Schematic of photocurrent response of an artificial optoelectronic synapse generated by laser pulses. f An optical image of the GNWs/CsPbBr 3 QDs synapse potentials in neurons based on stimulation, propagating through synapses for signal communication (Fig. 1b) [41]. Because of the ultra-high energy efficiency and the unique information processing mechanism, synapses conduct high-speed computing tasks with ultralow power consumption, which is almost infeasible in traditional computers. Each event of information transmission through the synaptic structure consumes only approximately 1-100 fJ [42]. The intensity of communication between neurons caused by neurotransmitters is defined as the synaptic weight, and the change in synaptic weight due to different external stimulus signals is called synaptic photoplasticity, which is considered to be the basis of brain learning and memory [43,44].
Artificial optoelectronic synapses were fabricated to perform human perceptual learning, in which the sensory ability can be enhanced through a learning experience (Fig. 1c). The structure of the optoelectronic synapse based on GNWs and CsPbBr 3 QDs is shown in Fig. 1d, and the fabrication flowchart and the optical images of the synapse array are illuminated in Fig. S1. First, GNWs were grown on a heavily p-doped Si wafer with a SiO 2 layer using a plasma-enhanced chemical vapor deposition (PECVD) equipment. Then, the GNW array was patterned by oxygen plasma etching. Next, positive and negative electrodes were fabricated by photolithography. Finally, the CsPbBr 3 QDs were spin-coated on the GNWs. The CsPbBr 3 QDs were spin-coated on the GNWs as light absorbers. The morphologies and structures of GNWs allow CsPbBr 3 QDs to be embedded within them, thus increasing the contact area with graphene.
In the optoelectronic synapse, gold electrodes can be regarded as axon terminals and dendrite terminals, GNWs/CsPbBr 3 QDs as vesicles, and photogenerated electrons/ holes of CsPbBr 3 QDs as neurotransmitters (Fig. 1b). When a pulsed laser is applied to the device, photocurrent changes can be induced (Fig. 1e). According to the optical image of the channel of the GNWs (Fig. 1f ), the length and width of the GNW area are 150 μm and 14 μm, respectively. The distance between two Au electrodes is 100 μm. Figure 2a shows a transmission electron microscopy (TEM) image of CsPbBr 3 QDs. The average diameter of the QDs is 13 nm. The detailed crystalline structures can be observed in the high-resolution transmission electron microscopy (HRTEM) image, as shown in Fig. 2b. GNWs exhibit a vertical nanosheet morphology with a sidewall height of 82 nm (Figs. 2c and S2). Through the Raman spectra of GNWs and GNWs/CsPbBr 3 QDs (Fig. 2d), a high-defect density of GNWs is shown, providing them with natural carrier-trapping characteristics [45]. Meanwhile, after doping GNWs/CsPbBr 3 QDs, the Raman shift of the G-peak changes from 1598.91 cm -1 to 1592.68 cm -1 due to the Kohn anomaly effect, indicating that CsPbBr 3 QDs and GNWs are n-type dopings [46,47].

Synapse characterization
Regarding the optical properties of the CsPbBr 3 QDs, we found that their absorption peak is mainly located at 500 nm. Steady-state photoluminescence (PL) of CsPbBr 3 QDs, shown in Fig. 2e, implies that under the excitation of a 450 nm laser beam the CsPbBr 3 quantum dot film emits a sharp emission peak centered at 515 nm, which is corresponding to the absorption spectrum shown in Fig. S3. When the CsPbBr 3 QDs were coated on GNWs, the PL intensity is weakened, or even quenched, and the emission peak is shifted, which indicates an efficient photogenerated carrier transfer caused by the GNWs/CsPbBr 3 QDs heterojunction. Through the GNW integration, the emission peak is red-shifted by 5 nm due to the narrowing of the bandgap of CsPbBr 3 QDs.
The PL time constant measurement of the CsPbBr 3 film and the GNWs/CsPbBr 3 QDs film was conducted under an excitation wavelength of 450 nm. The experimental results were fitted with a double exponential function to obtain the fluorescence lifetime parameters (Fig. 2f ). The fitting function: According to the average life expectancy calculation formula, where τ i is the PL time constant, and A i is the corresponding weighting factor. τ 1 and τ 2 are respectively related to nonradiative and radiative recombination [48]. The reference fitting function for the CsPbBr 3 QDs film can be extracted according to the following parameters: τ 1 = 21.92 ns, τ 2 = 182.83 ns, and A 1 = 0.6503, A 2 = 0.2525. The average life expectancy is 144.87 ns. On the other hand, the corresponding fitting function for the doped GNWs/CsPbBr 3 QDs film can be extracted according to the following parameters: τ 1 = 22.52 ns, τ 2 = 323.33 ns, and A 1 = 0.7296, A 2 = 0.2340. The average life expectancy is 269.66 ns. The results prove that the built-in electric field formed by GNWs and CsPbBr 3 QDs effectively separates the photogenerated carriers under laser beam excitation, thus reducing the recombination of photogenerated carriers in CsPbBr 3 QDs, resulting in decreasing of the fluorescence intensity and time constant, and increasing of photogenerated carrier lifetime. The extended carrier lifetime is crucial to achieving a (1)

Laser response characteristics of the synaptic device
The natural defects of GNWs can form the middle energy band (MGB) electron capture center on the boundary and surface (Fig. 3a). For the graphene with certain defects, the band gap opens [49]. The MGB refers to the defect structure of graphene, which forms a unique bandgap between the conduction and the valence bands [16]. GNWs have proved to contain many defects from the high D peak of the Raman spectrum (Fig. 2d). The photocurrent recovery time in GNWs is significantly longer than that of high-quality graphene films [50]. In high-quality graphene, the lifetime of photogenerated carriers is typically only picoseconds [51]. Therefore, it can be concluded that the defects in GNWs lead to the trapping effect of photogenerated carriers, which extends the lifetime of photogenerated carriers. The defects of GNWs themselves will cause the photogenerated carriers to be trapped, resulting in the slow recovery of the photocurrent after the laser beam was turned off. The emergence of the middle energy band reduces the electron-phonon scattering rate from the conduction to the valence band. Therefore, more photogenerated electrons and secondary generated electrons can be captured in MGB, resulting in a longer lifetime of photogenerated carriers [52]. After doping CsPbBr 3 QDs, a heterostructure is formed between GNWs and CsPbBr 3 QDs. The corresponding energy band diagram is shown in Fig. 3b. The direction of the built-in electric field points from the CsPbBr 3 QDs to GNWs. When illuminated by a laser beam, photogenerated carriers are mainly generated in CsPbBr 3 QDs, and GNWs also absorb the laser to generate some photogenerated carriers. The built-in electric field separates the photogenerated carriers so that the lifetime of the photogenerated carriers is further extended.
The photocurrent evolution of the HC-type optoelectronic synapses is measured. The photocurrent responses of the 530 nm laser pulse irradiation at different powers and exposure times are shown in Figs. 3c and S4. Prolonging the irradiation time and enhancing the laser power lead to an increase in the photocurrent intensity. This trend is not infinite, however, and the photocurrent tends to saturate after a period of illumination. A similar trend of the photocurrent changes is demonstrated under 630 nm and 450 nm laser pulse irradiation at different powers and exposure times (Fig. S5). The photocurrent response at different wavelengths is shown in Fig. 3d. The results under 0.35 W/cm 2 suggest that the lower wavelength can produce a stronger photocurrent, which is consistent with the changes induced by 0.70 W/cm 2 , 1.06 W/cm 2 , and 1.41 W/ cm 2 laser beams (Fig. S6). In the future, to reduce power consumption, one of the pathways is to decrease the light power. Moreover, reducing the device area is also effective, because the energy consumption is proportional to the light absorption area of the synapse. Through high-precision lithography technologies such as e-beam and two-photon absorption lithography, the energy consumption of the device can be greatly reduced, which can also effectively realize the fabrication of the device array. In addition, we suggest that reducing the contact resistance between electrodes and GNWs by optimizing the device fabrication process can also decrease energy consumption.
An excellent synaptic optoelectronic characteristic is shown under irradiation of a laser source with a wavelength of 360 nm and a power of 1.90 W/cm 2 (Fig. 3e). Herein the pulsed laser beam was repeatedly irradiated for 10 s with a 10 s interval, leading to a photocurrent change as high as 34.1 µA. The repeatability experiment was carried out at 1.90 W/cm 2 laser intensity, and the photocurrent intensity was maintained at 98.24 % at 400 s (Fig. S7). This indicates that under strong laser illumination, the CsPbBr 3 QDs degrade.
The device has significant voltage-modulated photoplasticity. As illustrated in Fig. 3f, the photocurrent changes can be suppressed by applying different gate voltages. Since the perovskite QDs are attached to the surface of the vertical-located GNWs, the electric field generated by the gate voltage (no matter whether upward or downward) weakens the interaction between QDs and n-type doped GNWs (Fig. S8). Consequently, the photocurrent response is reduced when the gate voltages are applied.
The GNWs/CsPbBr 3 QDs synapses exhibit significant long-term optoelectronic response. As shown in Fig. 3g, the half-life period of the photocurrent decay is 35.59 s after illumination by a 1.90 W/cm 2 laser pulse. Compared with the periods in reported devices (Table 1) [14,[20][21][22][23][24][25][26][27][28][29][30][31][32][33][34], the GNWs/CsPbBr 3 QDs synapses have both a long memory time and a microampere-level photocurrent, owing to the combined effect of the MGB and the built-in electric field formed at the GNWs/CsPbBr 3 QDs interface. A complete neuromorphic computing system requires the integration of the processor and many other electronic circuits. At this stage, electronic circuits typically work with microampere currents. Therefore, optoelectronic synapses that achieve microampere photocurrent responses can be integrated more effectively with other electronic circuits.
A controllable memory characteristic of the synapses is demonstrated through the paired-pulse facilitation (PPF) measurement. PPF refers to the enhanced postsynaptic response caused by two consecutive synaptic stimulations and exists in biological excitatory and inhibitory synapses. It plays a vital role in recognizing and decoding visual and auditory signals and other temporal-dependent information [53,54]. Figure 4a and b show the photocurrent curves of the synapses stimulated by laser pulses at different wavelengths. Using A 1 to represent the photocurrent change stimulated by the first laser pulse, and A 2 to represent the change stimulated by the second laser pulse, the PPF index refers to the ratio of A 2 /A 1 [55]. Figure 4c shows the relationship between the PPF index and the interval (Δt) of the two-laser pulses. The pulsed laser density is 1.90W/ cm 2 and the pulse frequency is 1 Hz. It is found that the relationship between A 2 /A 1 and Δt follows a double exponential function, where C 1 ( C 2 ) and t 1 ( t 2 ) are the initial facilitation degree and characteristic relaxation time of the decay term, respectively. t 1 of 0.24459 s and t 2 of 10.44171 s can be obtained by fitting. Obviously, t 2 is approximately one order of magnitude higher than t 1 , which is very consistent with the measurement data in biological synapses [56]. The photocurrent curves for the PPF measurement have been provided in Fig. S9. Moreover, the PPF behaviors also emerge on the synapses stimulated by 450 nm and 630 nm laser pulses (Fig. S10).
The memory duration of the GNWs/CsPbBr 3 QDs optoelectronic synapses can be controlled by the numbers and frequencies of laser pulses. Figure 4d and e show the photocurrent changes induced by different pulse numbers (1, 5, 20, and 50) or frequencies (0.1 Hz, 0.25 Hz, 0.5 Hz, and 1 Hz) of 530 nm laser pulses. When the pulse number and frequency increase, the photocurrent changes turn strong and the memory duration turns long. Pulsed laser stimulation at 450 nm and 630 nm also showed controllable (3) PPF = y 0 + C 1 * exp(−�t/t 1 ) + C 2 * exp(−�t/t 2 ) memory characteristics (Fig. S11), which are basic mechanisms of perceptual learning. An interesting phenomenon is that we found the photocurrent decays faster at 450 nm and 630 nm compared with the 530 nm laser. We suggest that it is because the photon energy at 530 nm wavelength is closer to the bandgap width of CsPbBr 3 QDs, so the built-in electric field prolongs the carrier lifetime more obviously.

Perceptual learning simulation
The GNWs/CsPbBr 3 QDs optoelectronic artificial synapse can mimic typical perceptual learning -an image recognition process of human retinal neuron networks. We measured the photocurrent of a GNWs/CsPbBr 3 QDs synapse was measured under different power densities and pulse numbers, and then simulated a 28×28 photocurrent array of the optical image of "8", which was selected from the Mixed National Institute of Standards and Technology (MNIST) dataset [57]. Herein random noises were added to the MNIST image (Fig. S12). Figure 5a presents the image of the digital "8" pattern induced by 1, 5, 10, 15, and 20 laser pulses with a power of 0.35 W/cm 2 . Under one laser pulse stimulation, the image resolution is poor owing to the influence of noise. With the increase in the pulse numbers, the signal-to-noise ratios gradually increase, and thus the images turn clear. After the laser pulses are applied, the further resolution change of the handwritten digit "8" can be considered a perceptual forgetting process. As shown in Fig. 5b, with the increase in the forgetting process duration, an image resolution decrease is demonstrated. This process is similar to the process of gradually forgetting after the brain remembers the image. Figure 5c shows the digital "8" images stimulated under various laser pulse powers. As the laser powers enhance the contrasts of the image increase because of the high photocurrent (Fig. 3c), indicating that the synaptic weights can be controlled using different powers. Next, to perform the perceptual learning with the MNIST handwritten digits, an ANN consisting of the 28 × 28 synaptic array and 10 output neurons was designed [58]. Figure 5d and e present the digital "8" grayscale image (28 × 28 pixels) and the ANN structure. The input is a 784-pixel image and the output neurons correspond to numbers 0-9. Figure 5f shows the image recognition result of 5 laser pulses. The percentage values of each category in the output neurons represent the probability of the corresponding number. Since the probability of the number 8 is the largest, the ANN successfully recognizes the handwritten digital image as the digital number "8. " Through the ANN we have demonstrated that the ability of sensory systems to respond to stimuli can be enhanced through a learning experience, which is a typical characteristic of perceptual learning. Figure 6a shows the recognition accuracies of 10000 MNIST test images corresponding to different pulse numbers and pulse powers. The accuracies are improved considerably with the increase of the pulse numbers. The accuracy after 20 laser pulses with a power density of 1.9 W/cm 2 is 96.42 %, around 2.4 times higher than that after one pulse. Moreover, the laser power increase can also improve the image recognition accuracy, but when the power density is above 1.06 W/cm 2 , the accuracies gradually reach saturation. We also realized a property of controllable perceptual forgetting, based on the long-term photocurrent response of the synapse. After training processes in which 5, 10, 15, and 20 1.90 W/cm 2 laser pulses are applied, the recognition accuracies under various forgetting durations were simulated. As shown in Fig. 6b, even after a 15 s forgetting process, the recognition accuracies can be well maintained above 90 %.
Combining the controllable properties of perceptual learning and forgetting, we have achieved a strong and continuous ability of image recognition under a low learning frequency. Herein the relationship between the pulse frequencies (0.1 Hz, 0.25 Hz, 0.5 Hz, and 1 Hz) and the recognition accuracies was investigated (Fig. 6c). High frequencies lead to high accuracies within seconds, while under a frequency as low as 0.1 Hz, a strong image recognition ability with accuracies above 80 % emerges after a 30-second training.
Another typical characteristic of perceptual learning, in which less time is required to regain the lost ability that has been previously learned, has also been realized. The GNWs/CsPbBr 3 QDs synaptic array was excited by two laser pulse trains with a power density of 1.90 W/cm 2 and a frequency of 1 Hz successively. As shown in Fig. 6d, the image recognition accuracies enhance by the first pulse train. It takes 17 seconds to enhance the accuracies from 93% to 97%. The confusion matrix of the 97 % accuracy after the learning is shown in Fig. 6e. At the next step, the accuracies decline during the train interval, which mimics the phenomenon that the learned images tend to be gradually forgotten. After the accuracies decrease to the level of 93 %, the second pulse train launches the relearning process of image recognition. It only takes 9 seconds for the second pulse train to recover the accuracy loss (Fig. 6f ). The interesting "learning-forgetting-relearning" behavior verifies that the mimicking of human perceptual learning can be conducted by the synaptic array with a long-term photocurrent response.

Discussion
Human perceptual learning which forms complex cognitive processes can be maintained throughout life. The recognition accuracy of human sensory systems is time-dependent, and can be determined by how many and when the related learning process was performed. In this paper, we have verified that the high accuracy of image recognition can be retained even under a low-frequency learning process. In perceptual learning, the learning process can differentiate the input signal of useful objects among a lot of information, and turn it effective to be recognized. With the accumulation of the learning experience, the ability to differentiate the information is gradually improved. Moreover, due to the memorial effect from the synapses in the neuron network, the relearning process takes less time and energy consumption than those in the first learning. It is a basic mechanism for a human to quickly and effectively perceive information through visual, auditory, tactile, olfactory, and taste.
In summary, we reported a HC-type GNWs/CsPbBr 3 QDs optoelectronic synapse for the demonstration of human perceptual learning. The synapses can be controllably stimulated by visible laser pulses at different numbers, powers, wavelengths, and frequencies. Through the introduction of GNWs as the QD substrate, the synapse has a long photocurrent decay time owing to the introduction of the unique MGB of GNWs and the use of the built-in electric field between GNWs and CsPbBr 3 QDs. Pulsed laser stimulation at 530 nm for 1 s, the half-life period of the photocurrent decay is as high as 35.59 s. The laser response intensity of the transistor-type structure design can be flexibly adjusted by applying an external gate voltage, which can imitate the regulation of the brain's synaptic excitement in the biological system. Moreover, the laser response results of the device can be processed in a typical neuromorphic way, such as PPF and memory duration tuning. The GNWs/CsPbBr 3 QDs optoelectronic synapse can be combined with an ANN, to mimic a kind of human perceptual learning -handwritten digit recognition. Under twenty laser power training, the recognition accuracy is as high as 96.42 %. A rapid accuracy increase is achieved by increasing the numbers and powers of the laser pulses, which is a typical characteristic of perceptual learning. In addition, we have shown that the relearning process takes less time to gain an accuracy enhancement than the first learning. The realization of human perceptual learning has great potential for high-performance artificial intelligence sensory devices.

Synapse fabrication
A heavily p-doped Si wafer with a 300 nm SiO 2 layer was rapidly heated to 910 °C at a rate of 200 °C/min and annealed for 5 min in an H 2 atmosphere with 500 sccm. During PECVD deposition of an GNW array, the pressure was controlled to be 5.5 mbar. 5 sccm CH 4 , 16 sccm H 2, and 300 sccm Ar were introduced into the chamber. After waiting for 15 s, the plasma was turned on for 12 min with a power of 30 W and a frequency of 10 kHz. Next, the GNW array was patterned by oxygen plasma etching and photolithography. 15 nm Ti/300 nm Au layers were thermal-deposited. Finally, the CsPbBr 3 QDs, bought from Suzhou Xingshuo Nanotech Co. Ltd, were spin-coated on the GNWs.

Synapse material characterization
The TEM images were acquired on a Tecnai F20 TEM. SEM measurements were carried out on a ZEISS Gemini 300 microscope. The CsPbBr 3 QDs solution was diluted at 1:50 and then 10 microliters were dropped on the silica glass slide. The absorption spectrum was measured by ultraviolet spectrophotometry with a wavelength range from 300 nm to 800 nm. A fluorescence spectrometer (Edinburgh FLS1000) was used to measure the emission spectrum and carrier lifetime of the CsPbBr 3 QDs. The excitation wavelength was 450 nm during the experiment, and the monitoring wavelength was 530 nm. We used HORIBA Raman equipment to measure the Raman spectra of GNWs. The laser wavelength selected was 530 nm.

Synapse characterization
All the electrical measurements were conducted with a precision semiconductor parameter analyzer (4200 SCS, Keithley) under a 4 V bias. The laser pulses were generated on a fiber-optic laser controlled by a signal generator. Figure S13 shows that the photocurrent curves in Figs. 3 and 4 were obtained by subtracting the dark current from the total current. Each group of experiment results was measured three times and taken with the average value. During the gate voltage measurement, a 4 V source-drain bias was used. We used 5 laser powers in our measurement (0.1 W, 0.2 W, 0.3 W, 0.4 W and 0.53 W), and the corresponding power densities are 0.35 W/cm 2 , 0.70 W/cm 2 , 1.06 W/cm 2 , and 1.41 W/cm 2 and 1.90 W/cm 2 .

ANN design and image recognition simulation
Twenty thousand training images and 10,000 test images of handwritten digits from 0 to 9, with 28 × 28 pixels, were selected from the MNIST dataset. Random noises were added to these images. The ANN contains 784 input neurons and 10 output neurons. The photocurrent of a GNWs/CsPbBr 3 QDs synapse was measured under different power densities and pulse numbers. We defined the photocurrent under the illumination of a 1.9 W/cm 2 beam with 20 pulses as the pixel grayscale value of 255. Next, a direct proportion relationship between the grayscale value and the photocurrent was set to obtain the photocurrent maps for image recognition. The simulation algorithm was developed by the Python language, in which the synaptic weights were updated by the stochastic gradient ascent/descent algorithm. The learning rate was determined to be 0.01, and the expected error value was set as 1×10 -3 . Each output neuron produced a signal from 0 to 1 using a sigmoid function.
Additional file 1: Fig. S1. The fabrication flowchart and optical images of a photoelectric synapse array based on GNWs and CsPbBr 3 QDs. Fig. S2. SEM images of the top-view GNWs. Fig. S3. An absorption spectrum of the CsPbBr 3 QDs. Fig. S4. Basic photoresponsive characteristics of the GNWs/CsPbBr 3 QDs optoelectronic synapse. Fig. S5. The photocurrent values of the synapses under different laser pulse powers and illuminating times. Fig. S6. Effects of different wavelengths and irradiation times on the photocurrent response. Fig. S7. Comparison of photocurrent intensity before and after 400 s illumination. Fig. S8. Transfer characteristic curve of the device. Fig. S9. The measurement results of the PPF property of the device with different laser pulse intervals. Fig. S10. The PPF effect of the synapse stimulated by laser pulses with different wavelengths. Fig. S11. The photocurrent of the device stimulated by 450 (a) and 630 (b) nm laser pulse trains under different pulse numbers. Fig. S12. Handwritten digital image processing for recognition simulation. Fig. S13. Total current measured under the laser density of 1.90 W/cm 2 .