Optical neuromorphic computing utilizes light-based neural networks for efficient AI processing. Conventional optoelectronic synapses are limited by single-dimensional perception and electrical weight modulation. This study overcomes these constraints by developing all-photonic artificial synapses using water-mediated phase transitions in MAPbI(3) perovskite. These synapses exhibit reversible light-driven optical memory capabilities, achieving a broad transmittance modulation via precisely controlled crystal deformation mechanisms. The high-stability synapses successfully replicate neurobiological functions, including paired-pulse facilitation, short-term to long-term memory transition, and humidity-dependent plasticity. Implemented within a recurrent neural network, the synapses achieve 100% classification accuracy for multidimensional optical stimuli encompassing power, duration, and environmental humidity parameters. Furthermore, integration with a diffractive deep neural network enables reconfigurable computing with 80 distinct programmable transmittance states, achieving remarkable classification accuracies on the MNIST handwritten digits and Fashion-MNIST datasets without hardware modifications. This work establishes a paradigm for developing intelligent systems that adapt to complex environmental changes, demonstrating potential for applications in dynamic visual perception and multi-task processing environments.

Two-terminal optoelectronic synapses have garnered great attention in neuromorphic computing due to their advantages in structural simplicity and energy efficiency. However, the device's operational flexibility is fundamentally limited by its spectral and electrical characteristics, which fail to tune photoresponses with wavelength selectivity and voltage modulation. Here, a MoS2/Au nanoparticle two-terminal optoelectronic synapse is presented that simultaneously achieves wavelength selectivity and voltage-modulated memory control. Adjusting the morphology of Au nanoparticles enables selective maximum photoresponses across the green and red wavelength regions, and voltage modulation induces switching between short-term and long-term memory states, thereby enhancing functional versatility. Based on this tunable photoresponse of the synapse, wavelength-selective image denoising and recognition tasks are demonstrated, achieving a maximum classification accuracy of 98.03% at 520 nm across the visible wavelength range. A wavelength-selective and voltage-modulated task for road condition classification is also implemented, incorporating an artificial neural network to track color-coded vehicles and classify their movement trajectories with precise directional recognition (96.67% accuracy). This work demonstrates a highly tunable optoelectronic synapse that achieves concurrent wavelength selectivity and voltage modulation in a single device, enabling neuromorphic computing and providing a practical strategy for developing highly efficient, multifunctional neuromorphic architectures.

Conventional optoelectronic synapses rely on electrical signals for core operations, resulting in complex circuitry, limited response speed, and energy inefficiency. Herein, an all-optical synapse based on perovskite MAPbBr2I is developed that directly converts optical stimuli into transmittance responses that mimic fundamental synaptic plasticity, including paired-pulse facilitation, short- and long-term memory, and learning. By using the dynamic transmittance response as input to an artificial neural network, high-accuracy dynamic pattern recognition of sequential characters is achieved. Furthermore, the optically controlled transmittance states are successfully integrated as programmable weights into a diffractive neural network, enabling all-optical classification of MNIST handwritten digits with an accuracy of 89%. This fully optical architecture, which eliminates electronic components and complex circuits, offers a promising pathway toward high-speed, energy-efficient vision systems by fundamentally circumventing the von Neumann bottleneck.

Breaking the optical diffraction limit is essential for advancing nanotechnology. This perspective discusses the advances of dual-beam nanoscale fabrication strategies, inspired by stimulated emission depletion microscopy, in enabling next-generation optical data storage and neuromorphic computing. It introduces the synergistic use of excitation and inhibition laser beams to enable super-resolution lithography, surpassing the diffraction limit to achieve feature sizes below 20 nm in photoresists and 90 nm in laser-scribed graphene. Key breakthroughs include petabit-capacity 3-dimensional optical disk memory using aggregation-induced emission materials and subdiffraction patterning of graphene for devices. Next, to achieve a breakthrough in computational power, the applications of these super-resolution techniques for fabricating optical neural networks are further outlined. We also highlight the integration of graphene-based optoelectronic synapses with optical neural networks, enabling nonvolatile information processing and paving the way for fully integrated optical neuromorphic architectures that combine sensing, storage, and processing with unprecedented energy efficiency.

Optical neuromorphic computing offers advantages such as high speed, parallel processing, and strong interference resistance compared with conventional electronic computing methods. However, electrical signals cannot be used as inputs to optical neuromorphic computing frameworks. Optical neuromorphic computing faces a tremendous challenge in converting electrical signals into optical responses. In this paper, a pathway for the conversion based on artificial electrochromic synapses is proposed. The synapse, comprising manganese dioxide and tungsten oxide, exhibits tunable optical transmittance under varying second-level voltage pulses. Based on the transmittance responses, typical behaviors, such as short-term/long-term memory transition and paired-pulse facilitation, can be observed. Moreover, image recognition based on artificial electrochromic synapses can mimic an adaptation effect to accelerate recognition speed and enhance energy efficiency. The responses of recognition accuracy changes turn weak gradually through voltage stimuli after repetitive exposures. The results imply that integrating artificial electrochromic synapses with optical neuromorphic computing offers a new approach to realizing low-cost, fast, and efficient artificial intelligence systems.

ABSTRACT In‐sensor computing integrates signal acquisition and processing within a single device, enabling real‐time, low‐power visual perception. Wavelength‐multiplexing holds great promise for leveraging the spectral dimension of light to expand parallel processing channels. However, current in‐sensor architectures typically rely on photodetectors with fixed spectral responses, limiting their ability to perform wavelength‐specific computations. Here, we present a neuromorphic wavelength photodetector integrated with 3D laser‐nanoprinted nanopillar metafilters that can simultaneously detect light intensity and wavelength across the visible range. The device features a bidirectional Indium Tin Oxide/n‐Si/Indium Tin Oxide heterojunction with two independently addressable regions, each covered by a metafilter with distinct transmittance spectra. By analyzing the photocurrent ratio under opposite biases, wavelength‐resolved detection is achieved with a resolution of 16.9 nm after accounting for the power‐dependent variation of the photocurrent ratio. Furthermore, the metafilters enable wavelength‐dependent binary weight encoding, thus we could demonstrate a binary neural network (BNN) model that switches computational tasks based on input wavelength—classifying handwritten digits at 520 nm and fashion items at 450 nm, respectively. This work provides a scalable and programmable approach for multifunctional, wavelength‐multiplexed in‐sensor computing, and highlights the potential of laser‐printed metasurface‐integrated photodetectors for future intelligent optoelectronics.

Reservoir computing bridges the gap between artificial intelligence and the physical environment. The realization of such systems depends on artificial synapses that emulate neurobiological plasticity in response to external stimuli. Functioning as reservoirs, these synapses convert the input signals into reservoir states, which are then retrieved by the readout layer. However, energy is merely dissipated during the operation of existing artificial synapses. Herein, we propose an energy-storage-involved reservoir architecture based on graphene-based supercapacitors. Reduced graphene oxide electrode, fabricated via metal-assisted microwave reduction, exhibits a mesoporous structure and low defect density. The supercapacitor delivers a specific capacitance of 352 F g-1 at a current density of 1 A g-1. Under current stimuli, the energy-storage device replicates biological synaptic functions through voltage responses. Next, the voltage responses can be used to mimic human perceptual learning, in which sensory systems are enhanced through experience. Moreover, integrated with a diffractive deep neural network in simulation, supercapacitor-based reservoir preprocessing can accurately identify static and dynamic handwritten images, outperforming an artificial neural network and an unprocessed diffractive deep neural network. These results pave the way for neuromorphic computing that integrates energy-storage-involved information processing, offering promising applications for energy-efficient artificial intelligence hardware.

Toward practical lithium-sulfur (Li-S) batteries, there is a pressing need to improve the rate performance and longevity of cells. Herein, we report developing a cathode electrocatalyst Lu SA/NC, capable of accelerating sulfur redox kinetics with a high specific capacity of 1391.8 mAh g- 1 at 0.1 C, and a low-capacity fading rate of 0.049 % per cycle over 1000 cycles even with a high sulfur loading (5.96 mg cm- 2). The unparalleled cathodes are built upon the unique structure in which single-atoms of rare earth metals are doped in nitrogen-doped porous carbon (RM SAs/NC). The theoretical and experimental studies reveal that the rare earth Lu atom has an unrivaled adsorption capacity for polysulfides and can promote facile deposition and dissolution reactions in charge-discharge processes. The in-situ Raman experiments provide direct evidence for its promotion of polysulfide transformation to eliminate the shuttle effect. The theoretical calculations suggest that the presence of f-dp hybridization enables accelerating sulfur reduction kinetics and enhancing lithium-sulfur battery performance. The strategic paradigm introduced in this study underscores significant practical potential in the exploration of rare earth single-atom catalysts for high performance Li-S batteries.

Developing artificial synapses capable of optical signal recognition is crucial for advancing neuromorphic computing. However, achieving high accuracy of synapse-based optical signal recognition, which avoids complex procedures of electrode fabrication and follows a contactless pathway, remains a significant challenge. In this study, we utilize a photochromic film of chemically synthesized WO3 with a transmission modulation of 77% to construct all-optical artificial synapses. Unlike optoelectronic approaches, the synapses leverage light for stimulation and contactless response measurement. Typical synaptic behaviors, including paired-pulse facilitation, learning experience, short-term memory, and long-term memory, can be demonstrated through transmittance responses under UV beam stimulation. Furthermore, the WO3 thin film can simulate the memory behavior of human skin's UV detection when transitioning from outdoor to indoor environments and back to outdoor conditions. Next, a recurrent neural network processes the synaptic transmittance responses to recognize optical signals, achieving 100% accuracy for preset light exposure durations and powers, 95% accuracy for closely spaced durations with a 1 s difference, and 100% accuracy for power differences of 14.5 mW. Moreover, 26 English alphabet letters encoded to different optical pulse trains can be recognized using an all-optical artificial synapse integrated with a recurrent neural network, achieving 100% accuracy. This work highlights the potential of photochromic materials in enabling high-performance neuromorphic computing and provides a new pathway for integrating optical signal processing with artificial intelligence.

Objective In the past few decades, artificial intelligence (AI) algorithms have been applied in various fields. Among them, neural network algorithms have become the common paradigm of modern AI and have achieved remarkable achievements in image recognition, natural language processing, speech recognition, and recommendation systems. However, the training process of these digital neural networks demands a large amount of time and energy. Therefore, optical computing solutions, with their multidimensional, high-speed, and low-energy advantages, have become a popular research area in AI applications. Extreme learning machine (ELM) is a machine learning paradigm where most connections in the model are established through randomly initialized nonlinear hidden nodes, and only a small part of the weights are adjusted during training after down-sampling. The advantage of this model is that it significantly reduces the training time as it replaces the time-consuming backpropagation with simple ridge regression. Nevertheless, in digital ELMs, the model performance heavily depends on the number of nodes in the hidden layer, which may lead to large memory consumption. To address this problem, we propose an optical extreme learning machine (optical ELM) by implementing random projections in the free-space propagation and investigating the effect of defocus on optical ELM. The optical aberrations, errors, and defects existing in the experimental process act as random components in the optical ELM, corresponding to the random transmission matrix in the digital ELM. This approach enables the passive realization of a large-scale hidden layer. We aim to design an easily deployable optical ELM that does not require complex processing of input and output data but achieves random projection through passive propagation in the optical domain. This method intends to simplify the system architecture while taking advantage of optical technology to achieve efficient parallel computing. Methods Fig. 1a shows the architecture of the digital ELM, defining the random projection process in the digital ELM as a randomly generated transmission matrix W, which describes the random linear mapping process between the input images and the output images. The matrix H is defined as the result of the input image matrix X multiplied by the transmission matrix W, with the ReLU function added as a nonlinear activation function. The calculation formula for H is given by H = ReLU(XW). Subsequently, the parameters are trained using the ridge regression algorithm with beta = ( HT H + cI )-1 HTT. Figs. 1(b) and 2 show the architecture and experimental setup of the optical ELM. A 532 nm wavelength laser is used, and its power is regulated by a Glan-Thompson polarizer and a half-wave plate. The input image is provided through a spatial light modulator (SLM) and propagates through free space. Optical aberrations, errors, and defects in the experimental process are defined as random transmission matrix W for the optical ELM. The matrix H is defined as the result of the input image matrix X multiplied by the transmission matrix W, using the nonlinear response function G of the camera as the nonlinear activation function. The calculation formula for H is given by H = G(XW), and the parameters are trained using the ridge regression algorithm with beta = ( HT H + cI )-1 HTT. Results and Discussions Figs. 3 and 4 show the simulation results for the digital ELM. First, under random projection, increasing both the input size and the number of hidden nodes remarkably improves ELM performance. Second, for high-resolution images, down-sampling in the optical domain is a more effective way to reduce computational burden. Figs. 5 and 6 illustrate the experimental results of the optical ELM. It shows that the utilization of the inherent random aberrations and errors of optical experiments and the nonlinear response of the camera in the framework is effective and can provide the necessary random mapping for optical ELM. Increasing the number of hidden nodes is related to the improvement of model performance. However, the propagation distance (PD) has a minimal impact on the model's performance. Conclusions We present a framework for optical ELM and provide a detailed analysis and experimentation. The experimental results show that the optical ELM can achieve a certain classification accuracy under specific parameter settings. By using the inherent random aberrations and defects during free-space propagation and the nonlinear response of the camera, this approach replaces the time-consuming and energy-intensive random mapping process used in digital ELM, thus enhancing hardware efficiency. This study validates the effectiveness of optical transmission in passively processing large-scale image data. Compared with other complex systems, this design only requires a simple deployment of optical pathways while achieving the initial research goal: to design an easily deployable optical ELM that does not require complex processing of input data. However, there is still room for improvement in this experiment. Thus, Fig. 7 shows a potential improvement scheme by introducing a nonlinear scattering medium during free-space propagation, which can provide the model with more randomness and nonlinear effects, thereby enhancing model performance.

Conventional electronics face intrinsic bandwidth and power constraints in deep learning, fueling the pursuit of optical computing's parallel processing and energy efficiency. A critical roadblock for optical neural networks (ONNs) lies in the missing trainable‐tunable nonlinear activation mechanisms required for modeling complex data correlations and cross‐task generalization. Here, this limitation is overcome through engineered tungsten trioxide (WO₃) thin films with dynamically controllable nonlinearity. Z‐scan measurements demonstrate stoichiometry‐dependent nonlinear responses modulated via photogenerated coloration, enabling precise control of optical nonlinearities. The films' non‐volatile memory effects permit implementation of adaptive in‐memory computing architectures where activation functions are trainable and task‐specifically optimized. This reconfigurable nonlinearity framework enhances both learning capability and generalization performance in ONNs, while achieving high‐performance THz‐scale response speeds (1 THz) for real‐time adaptive computing. Integrated as programmable activation layers in optical diffraction networks, the system demonstrates classification accuracy improvements of 4.62% (MNIST), 3.29% (Fashion‐MNIST), 13.53% (KMNIST), and 12.2% (CIFAR‐10). The synergistic combination of non‐volatile tunability, ultrafast reconfiguration, and stoichiometric control positions WO₃ thin films as a disruptive materials platform, resolving long‐standing activation function limitations in photonic neuromorphic systems while enabling adaptive high‐performance optical processors through on‐demand nonlinear engineering.

Neuromorphic computing systems convert multimodal signals to electrical responses for artificial intelligence recognition. Energy is consumed during both the response enhancement and depression, making the systems suffer from high energy consumption. This study presents a neuromorphic computing pathway based on supercapacitors. MXene Ti₃C₂Tx supercapacitors are fabricated and convert current stimuli to voltage responses. The response enhancement and depression are tunable through adjusting charging and discharging current stimuli, thus exhibiting synaptic plasticity. Typical synaptic behaviors are demonstrated, including short-term memory, long-term memory, paired-pulse facilitation, and learning experience. Next, the voltage responses are used to recognize Braille numbers represented by 3 × 4 arrays. A charging/discharging current pulse train representing each Braille array is applied to the supercapacitor. The voltage responses are collected and converted to 12-pixel greyscale images. Once the images representing Braille numbers 0-9 are input into artificial neural networks and deep diffraction neural networks, 100% accuracy can be achieved for recognizing the ten numbers. Because energy is stored during response enhancement in the supercapacitor and released once the response declines, this research demonstrates the potential applications of energy storage devices in neuromorphic computing, providing an innovative way to develop energy-efficient brain-like computing systems.

Light signals encode multidimensional parameters, such as wavelength, power density, and pulse duration, making visual perception critically important. The development of multidimensional light perception has proven to be a computational challenge. Conventional artificial visual systems consisting of optoelectronic sensors and von Neumann architecture suffer from separating sensors and memory units. Artificial optoelectronic synapses implementing optical memory have recently enabled neuromorphic computing for optical parameter classification. However, the classification of multiple light parameters on a synapse has not been achieved. Developing a synapse with adjustable photocurrent responses under multidimensional optical parameters and a neuromorphic computing paradigm suitable for the classification is crucial. Here, MoS2/SnO2 quantum dots optoelectronic synapses are demonstrated, in which the heterojunction between MoS2 and SnO2 achieves a pronounced optical memory effect. The time-dependent plasticity of the photocurrent responses is exhibited under wavelengths, power densities, and durations of light stimulation. The responses successfully emulate essential synaptic behaviors, including paired-pulse facilitation, short-term memory, long-term memory, and learning experience. Next, a recurrent neural network committed to processing the time-dependent responses is used to classify wavelengths, power densities, and durations of optical inputs. This realizes an accuracy of 100% under 1-parameter 3-class classification, 94% under 2-parameter 9-class classification and 96% under 3-parameter 8-class classification. Moreover, this work demonstrates effective feature recognition and extraction of a bicolor image, showcasing the advantage of multidimensional light perception for precise color-coded pattern extraction and advancing applications in multimodal image analysis. These findings highlight promising prospects in satisfying stringent performance requirements on artificial visual systems.

Recently optoelectronic synapses generating light-driven electrical memories have played a vital role in the neuromorphic computing of visual perception. However, all the optoelectronic synapses demonstrate photoelectric conversion. Peripheral circuits are used for contact photocurrent measurement, leading to significant energy consumption and impeding the evolution of optical wireless communication. It is crucial to develop noncontact neuromorphic visual perception based on light-driven photonic memories. Herein, we report all-photonic artificial synapses based on photochromic perovskites. Triggered by ultraviolet and visible light pulses, cesium lead iodide bromine induces a structural disorder. Optical transmittance changes induced by the disorder last after the pulses are gone. Next, the photonic memories are propagated in the air and processed by a recurrent neural network. The accuracy of binary image recognition is instantly stabilized at 1.0, and accuracy above 0.8 after 7 epochs is achieved in the recognition of digitals from 0 to 9. The all-photonic synapses realize remote perception with zero in-situ energy consumption and enable artificial sensory systems with low-power computation, remote control, and ultrahigh propagation speed. Optoelectronic synapses are key to artificial visual perception systems based on neuromorphic computing, but they typically rely on photoelectric conversion and peripheral circuits that are energy consuming and prevent optical wireless communication. Here, all-photonic artificial synapses with light-driven optical transmittance memories are fabricated based on photochromic CsPbIBr2 perovskite thin films.

The rapid development of neuromorphic computing has led to widespread investigation of artificial synapses. These synapses can perform parallel in-memory computing functions while transmitting signals, enabling low-energy and fast artificial intelligence. Robots are the most ideal endpoint for the application of artificial intelligence. In the human nervous system, there are different types of synapses for sensory input, allowing for signal preprocessing at the receiving end. Therefore, the development of anthropomorphic intelligent robots requires not only an artificial intelligence system as the brain but also the combination of multimodal artificial synapses for multisensory sensing, including visual, tactile, olfactory, auditory, and taste. This article reviews the working mechanisms of artificial synapses with different stimulation and response modalities, and presents their use in various neuromorphic tasks. We aim to provide researchers in this frontier field with a comprehensive understanding of multimodal artificial synapses.

Transferring the concept of chemical-driven responses into artificial intelligence technology holds the key to mimicking olfactory for neuromorphic computing of chemical recognition. Currently, artificial olfactory systems are designed based on chemical sensor arrays. Time-dependent responses of the sensor arrays are processed by artificial neural networks for recognition. However, the sensors generate instantly volatile responses, and algorithms for the processing of the time-dependent responses have not been involved. The recognition accuracy and speed are severely impeded. A sensor array can only achieve an accuracy of 90% after at least 5 training epochs. Herein an artificial olfactory chemical-resistant synapse consisting of 3D hierarchical WO3@WO3 nanofibers are demonstrated. The nanofibers exhibit persistent resistance responses through chemical exposures due to the strong chemisorption of water molecules. Typical synaptic behaviors including paired-pulse facilitation, long-term -1 short-term memory, and learning experience have been achieved. Next, a recurrent neural network that is committed to processing the time-dependent data is used to identify gas-phase chemicals of 3-hydroxy-2-butanone, triethylamine, and trimethylamine. Training-free gas recognition has been realized by a WO3@WO3 nanofiber synapse only, in which the accuracy is above 90% at the first epoch. The results have great potential to satisfy stringent performance requirements on artificial perception systems. An artificial olfactory chemical-resistant synapse consisting of 3D hierarchical WO3@WO3 nanofibers is demonstrated. The nanofibers exhibit persistent resistance responses through chemical exposures due to the strong chemisorption of water molecules and result in training-free gas recognition. A recurrent neural network is employed to identify gas-phase chemicals, achieving an accuracy of over 90% at the first epoch. image

The mimicking of human visual information processing, recognition, and storage is attracting intense interest in the field of artificial intelligence technologies. Electrochromic arrays, directly displaying images that can act as data sets for neuromorphic computing, are advantageous in providing a pathway to energy-efficient artificial visual perception. However, improvement of the recognition accuracy of low-contrast images is still a tremendous challenge. To establish a feature-enhancement strategy, a superlinear relationship between the responses and intensities of the input signals needs to be established. In this paper, reflective electrochromic arrays are fabricated by electrodeposition of Prussian blue and a bladed coating of carbon paste. The arrays exhibit a superlinear response of reflectance values at different voltage values. The reflectance almost remains stable in the range from 1.2 to -0.7 V and increases sharply below -0.7 V. The maximum reflectance modulation is as high as 74.8%. To enhance features of low-contrast digital images that are hardly recognized artificially, voltage values are generated proportionally from the grayscales of each pixel of the low-contrast images. Next, the electrochromic arrays display feature-enhanced digital images by controlling the voltage at each pixel. Consequently, artificial neural networks and diffractive neural networks take only 32 and 20 epochs to achieve 100% accuracy in low-contrast image recognition, respectively. The artificial visual perception design has great potential to realize sensory systems for pattern recognition from complex environments.

The von Neumann architecture, with its separation between the processing unit and memory, is struggling with limited performances and huge energy consumption burdens. To address the limitation, neuromorphic computing is emerging due to the combination of data storage and processing abilities. Exhibiting a light-driven persistence of electrical memory, optoelectronic synapses are crucial components to performing neuromorphic computing tasks. Herein, the typical behaviors of optoelectronic synapses such as short/long-term memory, paired-pulse facilitation, and learning experience are introduced. The development of various types of planar optoelectronic synapses, including perovskite-based synapses and two-dimensional oxide semiconductor-based synapses, is reviewed, and their applications in neuromorphic computing, including classical conditioning and image recognition, are described. Finally, the challenges in the field and the possible outlooks are discussed. Further explorations in this research area could focus on the investigation of optoelectronic conversion mechanisms and the nanofabrication techniques toward the development of all-photonic neuromorphic hardware.

The effects of thallium(Ⅰ) ions on the surface morphology, cathode current efficiency, cathode potential, polarization behavior, and electrochemical impedance spectroscopy of zinc electrowinning were studied by scanning electron microscopy and electrochemical measurements.The results showed that with increasing thallium(Ⅰ) ion concentration in the electrolyte, the hydrogen evolution reaction and the galvanic effect produced during zinc electrowinning increased.When the concentration of thallium(Ⅰ) ions in the electrolyte was 0.6 mg L -1 , the exchange current density of the zinc electrowinning process was maximum and the polarization was minimum.At this time, the Rct of the equivalent circuit was minimum, the CPE value was minimum, and the charge transfer rate was maximum.The cathodic current efficiency decreased from 80% to 55% when the thallium(Ⅰ) ion concentration was 1.5 mg L -1 .The presence of thallium(Ⅰ) ions also affected the surface macro-and microstructure of the zinc deposits.This result confirmed that thallium(Ⅰ) ions have a significant negative influence on the electrowinning of zinc.