Monolithic 2D Perovskites Enabled Artificial Photonic Synapses for Neuromorphic Vision Sensors

Neuromorphic visual sensors (NVS) based on photonic synapses hold a significant promise to emulate the human visual system. However, current photonic synapses rely on exquisite engineering of the complex heterogeneous interface to realize learning and memory functions, resulting in high fabrication cost, reduced reliability, high energy consumption and uncompact architecture, severely limiting the up‐scaled manufacture, and on‐chip integration. Here a photo‐memory fundamental based on ion‐exciton coupling is innovated to simplify synaptic structure and minimize energy consumption. Due to the intrinsic organic/inorganic interface within the crystal, the photodetector based on monolithic 2D perovskite exhibits a persistent photocurrent lasting about 90 s, enabling versatile synaptic functions. The electrical power consumption per synaptic event is estimated to be≈1.45 × 10−16 J, one order of magnitude lower than that in a natural biological system. Proof‐of‐concept image preprocessing using the neuromorphic vision sensors enabled by photonic synapse demonstrates 4 times enhancement of classification accuracy. Furthermore, getting rid of the artificial neural network, an expectation‐based thresholding model is put forward to mimic the human visual system for facial recognition. This conceptual device unveils a new mechanism to simplify synaptic structure, promising the transformation of the NVS and fostering the emergence of next generation neural networks.


Introduction
The explosive demand on data processing capability has placed ever-increasing pressure on traditional computers based on von Neumann architecture due to the physically separated memory and processor. [1]Human brain provides alternative inspirations for tackling this challenge.With ≈10 11 neurons connected via ≈10 15 synapses, human brain is a powerful processing system that can simultaneously store and compute information with extremely low energy consumption. [2]Information transmission enabled by a single synaptic event only consumes about 10 −15 J. [3] Synaptic plasticity enables the brain to efficiently process information and perform parallel operations.Neuromorphic computing has gained enormous attention for emulating the structure and parallel processing ability of the brain, showing high efficiency in image processing with high complexity.Recently, various types of neuromorphic devices have been proposed. [4]rtificial photonic synaptic devices, combining dual functions of photo-detecting and synaptic elements in a single device, have become promising candidates as they respond directly to external light stimuli, allowing temporary memory, and enabling real-time processing of optical information and data. [5]Thus, the photonic synapses can not only detect light signals but also track the history including light intensity, as well as the number, duration, and frequency of light spike to mimic visual perception of retina.Besides, compared to electronic synapses, photonic synapse devices also have wide bandwidth, high interference immunity, and low crosstalk characteristics. [6]o date, photonic synapses have seen significant advances on phototransistors by employing metal oxides, carbon nanotubes, organic polymers, and halide perovskites as the photoresponsive materials. [7]To emulate the memory of bio-synapse, the creation of a nonvolatile photocurrent after stopping the optical stimuli is essential for the photonic synapse.Typically, under an optical excitation, electron-hole pairs are generated while some of the charges are trapped at the interfaces between semiconductor/semiconductor, semiconductor/dielectric or semiconductor/charge transport layer due to the energy barrier or different electron/hole-withdrawing abilities.When the light is off, trapped charges are gradually released, enabling a persistent photocurrent.Thus, short-term potentiation, such as the excitatory postsynaptic current (EPSC) decay and paired-pulse facilitation (PPF) are imitated. [8]Similarly, artificial traps and semiconductor band alignment induced barrier also have been designed for storage and release of photogenerated carriers. [9]Thus, it is worth noting that most of the existing photonic synapses are developed by exquisitely designing and engineering complex interface and proper bandgap alignment to induce appropriate separation and storage of photogenerated carriers for photo-memory. [9]However, the extrinsic interface not only means complex and high cost in preparations, but also implies uncompact architecture and reduced reliability of the device, which are problematic for largescale fabrication and on chip integration.2D Ruddlesden-Popper perovskite with intrinsic organic/inorganic interfaces is a good choice to simplify synaptic structure, reduce fabrication cost, and minimize energy consumption. [10]Moreover, 2D perovskites that can be facilely obtained by solution methods exhibit relatively good stability, high degree of electronic tunability due to natural multiquantum-wells, and strong photo-response, making them suitable for photoelectronic devices.In this work, we developed a new photo-memory fundamental based on ion motion and ion-exciton coupling in 2D perovskite.A two terminal (2T) photo-memristor was fabricated based on monolithic 2D organic-inorganic hybrid perovskite (OIHP) to replicate synaptic behaviors under light stimuli.Our 2T photonic synapse without an extrinsic interface exhibits a highly neuron-like response to optical signals, allowing for adjustable synaptic plasticity and the emulation of learning and memory functions.All these intelligent properties make the device particularly suitable for emulating the human visual system.Image sensing, image memorization, and real-time preprocessing were demonstrated.The significantly enhanced performance also allowed us to design a single-layer synapse array for facial recognition without the use of complex artificial neural network.The device simulation exhibits an ultra-low electrical power consumption of 0.145 fJ when triggered by a light spike as weak as 0.07 mW cm −2 .This work developed a new platform to meet the complexity and freedoms required by neuromorphic vision sensors, which is of paramount importance for the next generation artificial photonics neural networks.

Retina-Inspired Neuromorphic Vision Sensor
The core elements in a biological visual system are retina, optic nerves, and the visual cortex (Figure 1a-i).In the retina, neurons are responsible for visual perception, which receives and converts external optical signals into electrical signals.These electronic signals are then converted to ionic ones, which are readout, processed and memorized by the synapses (Figure 1a-ii,iii).After being preprocessed by the retina, visual signals are then sent to the primary visual cortex for information processing and recognition.Inspired by the ionic transfer within biological synapse, artificial synapses have been developed based on voltage-induced ion flow. [11]As a mixed ionic-electronic conduction system, facile ion migration and illumination induced mobile ion within perovskites have stimulated interest in artificial synapses. [12]However, these existing demonstrations were still implemented through extrinsic interfaces.
Here, our conceptual design of an artificial retina consists of an optoelectronic synaptic array, with each synapse integrating functions of light perception, photo-to-electrical signal transduction and memorization (Figure 1b-i).The 2T synapses are constructed based on 2D OIHP single crystals, which consist of inorganic octahedra sandwiched between interdigitating bilayers of intercalated bulky alkylammonium cations. [13]It is believed that 2D OIHP is more stable with suppressed ion migration due to their higher formation energies and enhanced dielectric confinement by organic spacer cations. [14]Nevertheless, ionic motion has been frequently observed in 2D OIHP. [15]In particular, ion translocation within a single inorganic layer due to the strong electron-lattice coupling cannot be eliminated by the organic spacers.
As schemed in Figure 1b-ii, when a positive voltage is applied to the top electrode, an external electric field (E ext ) is created in the direction from the top to the bottom.Each electrical stimulation prompts iodine vacancy vertically migration through the inorganic layers, with negatively charged ions displacement against the direction of E ext .However, the large size of BA + is presumed to impede the migration of I − , thereby increasing the ion migration potential barrier.The ion redistribution is confined within individual inorganic layers.Thereby, positively charged vacancies swing to the upper inorganic layers, while negative ions build up at the bottom.As a result, a quantum confined polarization field (E pol ) is created within the inorganic layers.The polarization field can trap the photogenerated carriers at the organic/inorganic interfaces via Coulomb interactions.Note the ions only displace within the inorganic octahedron, which even may be activated by an octahedral distortion.The electrical activation energy is supposed to be evidently lower than that for mobile ion.In addition, illumination can further lower ionic activation barriers.Compared to the excited electrons, the activated ions possess much larger mass and size, much lower diffusion coefficient (≈10 −12 cm 2 s −1 ), [16] suggesting longer lifetime typically in the timescale of tens of seconds. [17]Therefore, the 2D layered OIHPs can produce persistent photoconductivity to replicate synaptic behaviors when being exposed to light stimuli (Figure 1b-iii).This conceptual design avoids the additional extrinsic interfaces and can be experimentally realized by one-pot chemical synthesis, thus significantly simplify the device structure and fabrication process.

Nonvolatile Photocurrent in 2D Layered (BA) 2 PbI 4 Crystals
The 2D layered OIHP (BA) 2 PbI 4 (n = 1) crystals were grown using a conventional slow cooling method (see photos in Figure S1, Supporting Information).The low cooling rate during growth minimizes local lattice distortion. [18]X-ray diffraction (XRD) measurements were performed to investigate the structure of (BA) 2 PbI 4 single crystals.Figure 2a shows evenly spaced reflections below 2 = 14°, indicating the formation of pure-phase homologous 2D layered single crystals.The sharp and well-resolved diffraction peaks corresponding to (00l) (l = 2, 4, 6…) lattice planes indicate high crystallinity and orientation. [19]Top-view optical and scanning electron microscope (SEM) images of the (BA) 2 PbI 4 single crystal reveal a smooth surface and plate-like crystal morphology (Figure S2, Supporting Information).
Figure 2b displays the optical absorption and photoluminescence (PL) spectra of the (BA) 2 PbI 4 crystals.The light absorption shows an edge at 2.38 eV with a broad while weak band at about 700 nm.Dual emission peaks at 521 nm (2.38 eV) and 564 nm (2.20 eV) are observed.According to previous reports, [20] the PL peak at 564 nm is assigned to the excitonic emission in bulk undergoing a photon recycling while that at 521 nm is ascribed to the emission at the surface (Figure S3, Supporting Information).Transient absorption spectra contain a sharp photobleaching (PB) band at 525 nm and a broad PB band at 700 nm (Figure 2c and Figure S4, Supporting Information).Evidently, the photobleaching (PB) band at 525 nm corresponds to the excitonic recombination.Broad PB band at 700 nm is consistent with the steady absorption shown in Figure 2b.Previous reports have verified the broad band transition is related to the halide vacancies. [21]Compared to 3D perovskites, the suppressed ion migration in 2D layered perovskites is explained by an increase in the energy required to form an ion vacancy. [22]Here, the preexistence of halide vacancies during crystal growth offers the channels for ion migration.
Figure 2d and Figure S5 (Supporting Information) show the repetitive photo-switching characteristics of the 2T ITO/(BA) 2 PbI 4 /ITO photodetector at different wavelengths when subjected to light pulses with a repetition rate of 0.2 Hz.Notably, a significant photo-response is observed at the wavelength range of 500-550 nm, while only a weak photocurrent is generated when stimulated with light pulses of 700-800 nm under the same conditions.The responsivity (Res = I light /P light ) of the photodetector to light pulses at different wavelengths is depicted in Figure 2e.The responsivity at the wavelength range of 500-550 nm is 68.93 mA W −1 , which is 1.50 and 4.46 times higher than those at 400-450 and 600-650 nm, respectively.2D layered OIHP single crystals have been confirmed to show significantly improved stability, greatly beneficial for applications. [23]o evident degradation was observed over repeated switching cycles (Figure S6, Supporting Information).Figure 2f shows a typical on/off switching of the photodetector under a 410 nm light pulse (2 mW cm −2 ) at a bias voltage of 0.1 V.The photocurrent sharply increases upon switching on the light pulse with a rising time of ≈7 s.Notably, after removing the light pulse stimulation, the photocurrent stimulated by the optical pulse persists for an extended period, with a fall time (defined as the time necessary to reduce the photocurrent from 90% to 10%) as long as 91.52 s.This persistent photocurrent of the vertical 2T photodetector to temporarily store optical information is not observed in the horizontal 2T photoconductors based on (BA) 2 PbI 4 crystals (Figure S7, Supporting Information).Only in the vertical configuration, the applied bias voltage induced polarization can significantly promote the formation of exciton-ion complexes at the organic-inorganic interfaces (Figure 1b-ii).
To gain insight into the nonvolatile photocurrent, electrical characterization was performed by grounding the bottom ITO electrode and sweeping the voltage on the top ITO electrode.Figure 2g shows the resistive switching effect of the ITO/(BA) 2 PbI 4 /ITO photodetectors under a direct current (DC) voltage sweep with different stop voltages (0 V → 0.1 ≈ 0.5 V → 0 V).The stimulated current exhibits a nonlinear increase with voltage.Moreover, bidirectional resistance switching characteristics with loops are observed.The currents in reverse scanning are smaller than those in forward scanning.Similar bidirectional resistance switching characteristics with loops are also observed in negative bias voltage (Figure S8a, Supporting Information).Under the same stop voltage and scanning rates, the hysteresis between forward and reverse scans decreases under constant light illuminations of 0.5 mW cm −2 (Figure S8b, Supporting Information).As compared in Figure S8c (Supporting Information), the photo-current is greater than that in the dark due to the photoresponse.Regardless of dark or light illuminations, the I-V hysteresis shows excellent stability and repeatability, implying a reliable property of the 2D OIHP (BA) 2 PbI 4 crystals, which is similar to the ferroelectricity that can maintain a nonvolatile state.In fact, I-V hysteresis in the perovskite solar cells has been extensively observed, which is commonly explained by the slow-moving ions. [24]The observation of I-V hysteresis confirms the ion motion in 2D OIHP.
In addition, the PL spectra were measured when multiple light pulses (10 light pulses of 1 Hz, 0.8 mW cm −2 ) were applied continuously under applied bias voltage.As shown in Figure 2h, the PL intensity at 521 nm increases steadily with the number of applied light pulses (inset in Figure 2h).The PL intensity measured under the second light stimulation can be enhanced by the residual photogenerated carriers excited by the first light stimulation, which is consistent with the nonvolatile photocurrent.TRPL decay curves of (BA) 2 PbI 4 crystal under multiple light pulses with fixed bias voltage are shown in Figure 2i, and the curves can be well fitted in terms of biexponential decay.20a] With increasing number of applied light pulses, the PL lifetimes increase from 0.47 to 0.70 ns, indicating increased bulk emission.Besides, the PL spectra under multiple continuous excitation pulses without bias voltage are shown in Figure S9 (Supporting Information).The continuous spike induced PL enhancement is much weaker due to the absence of polarization.It is worth noting that multiple excitation pulse induced PL enhancement is different from the light-healing effect reported previously (Figure S10, Supporting Information). [25]erein, the nonvolatile photocurrent, I-V hysteresis, and the multiple excitation pulses induced PL enhancement are all related to the mixed ionic-electronic conduction in 2D OIHP.The strong interactions between exciton and ion motion induced polarization field suggest the possibility of larger excitonic complexes.Analogous to the well-known trions, the radiative transitions of the exciton-ion coupling system are not allowed by the selection rules. [26]Thus, the excitons cannot recombine until the exciton-ion de-coupling.Consequently, the excitons are stored for a timescale of seconds. [27]After removal of the light stimulus, the gradually recovery of exciton due to exciton-ion de-coupling is responsible for the nonvolatile photocurrent and residual photogenerated carriers.More discussion on the exciton-ion coupling model can be found in Figures S11-S13 (Supporting Information).

Light Pulses Modulated Synaptic Characteristics
We demonstrated that the vertical 2T photodetector based on the (BA) 2 PbI 4 crystal (Figure 1b-i) is able to emulate key characteristics of biological synapses.In a biological neural network, a synapse is the interconnection between the axon of a preneuron and a post-neuron's dendrite, functioning as a transmitting channel to convert a chemical signal into an electrical one (Figure 1a-ii).Pre-synaptic neurons contain neurotransmitters and release them into the synaptic cleft after receiving an impulse.Post-synaptic neuron receptors accept the diffused neurotransmitter, resulting in changes in potential and generation of post-synaptic current (PSC). [28]The ability to modify the connection strength, i.e., the synaptic weight, is known as synaptic plasticity, which is responsible for the complex learning process in the human brain. [29]ere, a continuous voltage bias is applied to the 2T memristor as a reading voltage, which is small enough (0.1 V) to minimize its effect on the device current.Analogous to presynaptic spikes, light pulses at 410 nm are applied while measuring photo-stimulated current between the 2T photo-memristor, termed the synaptic weight.The photo-stimulated current level increases abruptly after one optical pulse with a duration of 3 s, which is defined as EPSC, indicating that optical stimulation contributes to exciting the connected post-neuron.This non-volatile current can be stepwise erased by the application of reverse electrical pulses (blue regions), which demonstrates the capability of photonic writing and electrical erasing of the 2T photodetector (Figure 3a).This optical response behavior in the vertical 2T photodetector is analogous to the short-term plasticity (STP), which is an important temporal manifestation of short-term memory (STM) in biological synapses.STP is involved in encoding temporal information in auditory and visual signals, which plays a vital role in associative learning, pattern recognition, information processing, and sound source localization.Paired-pulse facilitation (PPF), as an essential characteristics of STP, is considered as the basic operation of temporal information encoding of visual signals, describing the temporary enhancement ability of synaptic weight facing two successive stimulation signals. [6]s shown in Figure 3b, during the stimulation of two consecutive light spikes with an interval time (ΔT) of 3 s, the EPSC triggered by the second light spike is larger than that triggered by the first one.The device obtains a more prominent EPSC peak based on incomplete recombination charges induced by the first spike when the second light stimulation occurs.4a] The dependence of the PPF index on ΔT is shown in Figure 3c, where the light pulse intensity is fixed at 0.5 mW cm −2 , and the illumination duration is fixed at 3 s.As ΔT decreases from 30 to 1 s, the PPF index gradually increases from 109% to 126%, exhibiting an exponential relationship.These results demonstrate that the synaptic device based on the 2D perovskite can indeed mimic the STP of a biological synapse.
In biological synapses, the formation of synaptic weights in long-term plasticity (LTP) can last for tens of minutes or even years.The transition from STP to LTP can be achieved by ongoing or repeating stimuli that persistently enhance the synaptic strength and store information over a long period of time.To further examine the potential LTP of the artificial synapses, we investigated the synaptic weights using various light pulse modulations, including light pulse number (N light ), illumination pulse duration (T light ), and illumination intensity (P light ).As shown in Figure 3d, the EPSC increases steadily when the number of optical spikes (T light = 3 s and ΔT = 2 s) increases from 1 to 5, mimicking the spike-number-dependent potentiation (SNDP) of brain learning process.After a 5-pulse stimulation, a longsustained current is observed, contributing to a largely increased relaxation time of 13.579 s (the time taken for the EPSC to drop to 30%).Furthermore, as shown in Figure 3e,f, the EPSC can be efficiently altered by the light pulse's illumination time and intensity, corresponding to spike-time-dependent potentiation (STDP) and spike-intensity-dependent potentiation (SIDP).These transitions from STP to LTP can be attributed to the larger amount of photogenerated charge carriers under a longer duration time, shorter interval time, and higher intensity.As depicted in Figure 3g,h, light spikes at different frequencies can also modulate synaptic weights.When the spike number keeps constant, the EPSC relaxation time becomes shorter with an increase in frequency, which is attributed to the shorter effective illumination time at higher spike frequencies, in consistent with the STDP.
When human study new knowledge, the brain will experience three processes including learning, forgetting, and relearning.The relearning process requires less time to restore the forgotten information than the initial learning process.Here, this learning process in the human brain is emulated using the vertical 2T photodetector based on the (BA) 2 PbI 4 crystal.As shown in Figure 3i, first, 22 consecutive light pulses (T light = 0.7 s and ΔT = 0.3 s) were used to excite the photodetector.Then, the photocurrent spontaneously decreased to an intermediate state after removing the light stimulation.Subsequently, only 7 consecutive light pulses are needed to recover the device to the previous learning level, much fewer than the number of optical pulses needed in the first learning process.When forgotten again for 7 s, only 5 pulses are needed for the next learning.Besides the simulation of the biological synaptic functions, low energy consumption is another important feature for an ideal synaptic device to enable large number of network interactions for processing complex information.We demonstrated the obvious synaptic behavior that could be still observed when applying a low V bias of 1 μV (Figure S14, Supporting Information).The electrical power consumption per synaptic event is estimated to be as low as 0.145 fJ, which is one order of magnitude lower than the typical energy consumption of biological synapses (Table S1, Supporting Information). [30]A comparison of synaptic characteristics with other photonic synapses in the literature is shown in Table S2 (Supporting Information).
The photodetectors based on (BA) 2 PbI 4 crystals with different sizes and thicknesses always exhibit similar I-V hysteresis and synaptic characteristics (Figures S15 and S16, Supporting Information), indicating miniaturization of the device for on-chip integration is applicable.In addition, when the bottom ITO is replaced by a Cu film, similar I-V hysteresis and synaptic characteristics still present (Figure S17, Supporting Information), further verifying that these characteristics are attributed to ion motion within the (BA) 2 PbI 4 crystal, rather than the interfacial contact between the (BA) 2 PbI 4 crystal and ITO.Moreover, the 2D perovskites with higher n are also feasible for photonic synapses (Figure S18, Supporting Information).

Image Recognition with Preprocessing by the Photonic Synapse
In the field of artificial neural networks (ANN) for pattern recognition, the image processing chain comprises five distinct tasks including image preprocessing, data reduction, segmentation, object recognition, and image understanding.Exploiting the contrast enhancement capacity of image preprocessing can effectively draw attention to the main features of the image and extract corresponding features for classification.Figure 4a depicts the schematics of the human visual system.The visual information is first detected and extracted through the retina in the human eye.The information is further passed through optic nerves and processed in the visual cortex.Our photonic synaptic device is constructed to receive presynaptic spikes and temporarily store information.The memory time is influenced by different optical stimuli, which can be leveraged to facilitate image preprocessing.Specifically, we developed a system that includes image sensing, preprocessing, and recognition based on the synap-tic arrays (Figure 4b).We selected images from the widely used MNIST (Modified National Institute of Standards and Technology) dataset for training and testing.A synaptic array consisting of 28 × 28 units (Figure S19, Supporting Information) was designed to receive external optical signals, which was employed as the preprocessing component to extract feature information from the images.The preprocessed images were then supplied to a 784 × 30 × 10 three-layered ANN for subsequent training and recognition.Here, the 784 neurons in the input layer corresponds to each pixel of the input image, while the output layer comprises 10 neurons, corresponding to the classification tags of 0-9.
For the demonstration, we randomly selected three image examples from a database containing 600 images.Background noise signals in each image were randomly generated.Supposing the images were illuminated by a 410 nm pulse, each synaptic unit was stimulated at 410 nm.The pixels related to numerals mean strong light stimuli (1.2 mW cm −2 ) to the synaptic unit while the noise pixels correspond to relatively weaker (0.07-0.32 mW cm −2 ) light stimuli.According to the SIDP of the photonic synapse, noise signals decay faster than that of the numerals (Figure S20, Supporting Information).Figure 4c shows a comparison of the input and output preprocessing images at different decay times.All the EPSC signals in the image were normalized.After the preprocessing through the synaptic array, the handwritten numerals were highlighted while the background noises were smoothed.Subsequently, the preprocessed images with enhanced contrast were put into the ANN to conduct the image training and recognition.Figure 4d shows the significant improvement in the recognition accuracy achieved through image preprocessing using our synaptic array.Without the image preprocessing by synaptic array, only 18 handwritten numerals were correctly identified in 100 test samples.However, with the preprocessing by our synaptic array, the recognition ability and correlation coefficient were gradually improved.At decay time of 50 s, the recognition accuracy reached 74%.By increasing training epochs, the recognition rate reaches 94.01%.The improved recognition ability is attributed to the decreased noise and error evaluation values.More details, such as recognition results for noisy images with and without preprocessing by synapse, are presented in Figures S21,S22, and Table S3 (Supporting Information).In addition, Figure S23 and Table S4 (Supporting Information) also compare the recognition results for noisy images without preprocessing and unnoisy images.These results suggest that our synaptic device has the capacity for image preprocessing through both contrast enhancement and noise reduction, commonly known as feature extraction in convolutional neural networks (CNNs).

Artificial Retina for Face Recognition Based on a Self-Constructed Model of Synaptic Array
Based on the above experimental results of image recognition, we further put forward an expectation-based thresholding model to mimic the human visual system for facial recognition without the use of an ANN.First, based on the predictive capability of transient currents of the devices under a specific illumination condition (Figure S24, Supporting Information), the synaptic array was trained with nine grayscale face images of a man with a distinct facial expression and orientation.The facial regions with stronger light reflection led to higher memory currents.Therefore, a memory current mapping that retained the facial characters was generated (Figure 5a).Then, a subset of the mappings, which not only had a high memory current, but also could represent the facial characteristic outline, was used for building the facial recognition model (Figure 5b).We adaptively set a decisionmaking threshold for each synaptic unit by using this model (Figure 5c).When the synaptic array senses a face that the same as the training one, most of the synaptic units retaining high memory current are further activated, causing a photo-current exceeding the decision threshold due to the PPF.Otherwise, the memory current mapping obtained from training mismatches with the testing face, resulting in an idle status.
As a demo, four example faces were tested by the face recognition model based on our visual sensors.As shown in Figure 5d, when facial images of other men were received, the activation rate of the artificial retina was smaller than 70%.In contrast, the artificial retina showed an activation rate of 89.91% when testing with the target man's face.Even when the target man's face contained 30% noisy pixels or from different angles and different expressions (Figure S25, Supporting Information), the artificial retina still showed more than 80% of activation rates.These results suggest that the artificial retina can effectively learn critical visual features for recognizing target images.Moreover, by taking advantages of the spectral response and synaptic plasticity, the artificial retina is a filter-free color-cognitive device (Figure S26 and Table S5, Supporting Information).

Conclusion
In summary, we proposed a new persistent photocurrent mechanism based on ion motion and ion-exciton coupling in 2D perovskites.Due to the intrinsic organic/inorganic interface within the 2D layered OIHP and the strong exciton-ion coupling, the photodetector without an extrinsic interface exhibited a persistent photocurrent lasting about 90 s after removing the illumination.The special charge carrier dynamics in (BA) 2 PbI 4 crystals exerted a significant impact on the optical memory function.The 2T

Experimental Section
Synthesis of (BA) 2 PbI 4 Single Crystals: The raw materials, PbO, 57% aqueous hydrogen iodide (HI), H 3 PO 2 and C 4 H 9 NH 2 (BA) were used for chemical reactions.PbO (10 mmol) was dissolved in a mixture of HI (61 mmol)/H 3 PO 2 (12 mmol) and heated to 90 °C under continuous stirring to obtain a bright clear yellow solution.Then, the mixture was annealed at 90 °C for 5 min.In a separate beaker, BA (4 mmol) was neutralized with HI (4 mmol) in an ice bath.The cold neutralized C 4 H 9 NH 3 I (BAI) solution was added dropwise into the hot PbI 2 solution with continuous stirring.The temperature drops rapidly when the solution was left to cool at room temperature, many small orange flakes of the 2D (BA) 2 PbI 4 perovskites were formed.In order to grow high-quality millimeter-sized monolithic crystals, the solution is still left to cool in the closed heating unit which can greatly slow down the cooling rate.Orange (BA) 2 PbI 4 crystals were obtained by drying for 1 h in a vacuum oven (55 °C) to remove the surface adsorbed solvent.
Fabrication of Synaptic Photodetector: Glass substrates coated with ITO (ITO-glass) were cleaned ultrasonically in acetone, ethanol, and deionized water for 20 min each and dried by using N 2 gas to prevent the substrates from being contaminated by impurities in the air, then treated by UV-zone for 20 min to increase wettability.Then, the substrates were subjected to plasma treatment for 6 min.Subsequently, a dried single crystal of 2D perovskite was placed on the ITO-glass substrate.Several layers of scotch tape were attached around the 2D perovskite to insulate the two electrodes.Finally, ITO/(BA) 2 PbI 4 /ITO was obtained by attaching another piece of ITO-glass to the tape.The two electrodes were compressed tightly by a clip.The top view photo and cross-section image of the device taken through a microscope are shown in Figure S27 (Supporting Information).In addition, Cu/(BA) 2 PbI 4 /ITO devices were obtained by depositing Cu electrodes on glass substrates by e-beam evaporation.
Characterizations: PL spectra were obtained by a Maya 2000 Pro highsensitivity spectrometer (Ocean optics) equipped on an optical microscope, and a 405-nm continuous-wave laser was used as the excitation source.Time-resolved photoluminescence (TRPL) spectra were measured on a time-correlated single photon counting setup (TCSPC, TimeHarp 260, PicoQuant) using a 375 nm ps pulsed laser as the excitation source.The UV-vis optical absorption spectra were acquired with a Shimadzu UV2600 spectrometer with an integrating sphere.D/max 2500/PC rotating target X-ray diffractometer was used to measure the X-ray powder diffraction (XRD) patterns.SEM images were captured by using a Thermo Scientific field emission scanning electron microscope (FE-SEM, HitachiS-3000N, Japan).The I-V curves were measured using the Keithley 2450 semiconductor characterization system.

Figure 1 .
Figure1.The concept of an artificial visual system mimicking the biological system.a) The biological visual system consisting of the retina (receiving and preprocessing), optic nerves (transmitting), and the visual center (processing and memory system) and a multilayer structure of a retina.b) The artificial visual system based on a 2T vertical photodetector of ITO/(BA) 2 PbI 4 /ITO.The exciton-ion coupling is responsible for the nonvolatile photocurrent.

Figure 2 .
Figure 2. Structural characterizations and photo-response.a) X-ray diffraction pattern of the (BA) 2 PbI 4 crystals.b) Absorption (blue) and PL (orange) spectra of the (BA) 2 PbI 4 crystals.c) Transient absorption spectra of the (BA) 2 PbI 4 crystals.d) Photo-switching stimulated by light pulses at different wavelengths (0.07 mW cm −2 and V bias = 0.1 V). e) Normalized responsivity of the photodetector at different wavelengths.f) Photocurrent response under a single light pulse lasting 20 s (2 mW cm −2 and V bias = 0.1 V). g) I-V curves under a direct current (DC) voltage sweep with different stop voltages (0 V → 0.1 ≈ 0.5 V → 0 V).h) PL spectra of (BA) 2 PbI 4 crystal when multiple excitation pulses are applied continuously under applied bias voltage.Inset: PL intensities at 521 nm versus the number of applied excitation pulses.i) TRPL decay curves of (BA) 2 PbI 4 crystal under multiple light pulses with fixed applied bias voltage.Inset: average PL lifetime obtained by bi-exponential fitting versus the number of applied excitation pulses.

Figure 3 .
Figure 3. Mimicking the synaptic characteristics.a) Characteristic photonic responses of the vertical 2T photodetector with light writing (green region) and voltage erasing (blue regions).b) PPF behavior at two successively applied light pulses (T light = 3 s).c) Dependence of the PPF index on the ΔT, which is fitted by an exponential function.Transition from STP to LTP by d) using the repeated light pulses (ΔT = 3 s, P light = 5 mW cm −2 ), e) lengthening the T light and f) enhancing P light .g) Transition from STP to LTP by reducing frequency under the same effective illumination time.h) Transition from STP to LTP by reducing frequency under the same number of light pulses.i) Learning experience including learning, forgetting, and relearning processes under 410 nm light pulses (2 mW cm −2 ) with a pulse frequency of 1 Hz.

Figure 4 .
Figure 4. Image recognition with preprocessing based on the photonic synapse.a) Schematics of the human visual system.b) Illustration of the synaptic array for image preprocessing and an artificial neural network for image recognition.c) The input and output images at different preprocessing times.d) At decay time of 50 s, the recognition accuracy at different epochs.Inset: Image recognition accuracy for noisy images with and without preprocessing by the photonic synapse.

Figure 5 .
Figure 5. Artificial retina for face recognition.a) Illustration of the model training of the artificial retina.b) A subset of the synaptic array was selected for monitoring facial recognition.c) Decision-making condition for classifying active and idle status of each unit.d) Facial recognition results based on the training model without ANN.Credit: photographs of faces, AT&T Laboratories Cambridge, data from ref. [31].