Complex interactions between neurons can be simulated using multi-layer networks, such as cross-layer coupling between neurons in different brain regions. This study employs numerical calculations to investigate the synchronization characteristics of multi-layer optical pulse neural networks based on vertical-cavity surface-emitting lasers with saturable absorber (VCSEL-SA). Considering the limited signal transmission speed in neural systems, we comprehensively evaluate the impact of intra-layer and inter-layer delays on network synchronization. Our findings reveal that different coupling delays effectively induce transitions in network synchronization patterns. Furthermore, we examine the influence of key VCSEL-SA parameters on synchronization stability and demonstrate the robustness of neuronal synchronization against parameter mismatches between different layers. Finally, we validate the universality of our conclusions through a three-layer photonic neuron network. This work presents a systematic and in-depth investigation of synchronization characteristics in multi-layer networks composed of photonic neurons. The results provide valuable insights for practical applications of brain-inspired optical neural networks.

Based on the laser-induced breakdown spectroscopy assisted with laser-induced fluorescence (LIBS-LIF) technology, this paper proposes a feature equilibrium method to fuse the feature data of homologous laser-induced breakdown spectroscopy (LIBS) to analyze the lead content in kelp with high precision. First, LIBS-LIF technology is employed for univariate quantitative analysis of lead in kelp, with a detection limit of 0.085 mg/kg, which is well below the limit set by national standard, and the sensitivity meets the detection requirements. Then, the LIBS-LIF spectra are interpolated using cubic spline interpolation, while the peak fragment selection for the LIBS spectra is performed using a genetic algorithm. This process resulted in a feature-equilibrium spectrum with matching feature dimensions. Finally, the proposed method is validated using three machine learning algorithms: ridge regression (RR), random forest regression (RFR), and support vector regression (SVR). Results show that the feature equilibrium method significantly improved the prediction accuracy and generalization ability of the model, among which the SVR model achieves the best quantitative performance, with the coefficient of determination, root mean square error, and average relative error for prediction set are 0.957, 0.251 mg/kg, and 7.36%, respectively. It is proved that the proposed method offers high quantitative accuracy, which provides a new approach for achieving high-sensitivity and high-precision detection of lead in kelp.

To address the low accuracy of PCBA defect detection caused by the lack of 3D morphological information, an MGT-Fusion defect detection method incorporating both RGB texture and depth image features is proposed. The proposed method enhances traditional RGB texture image-based defect detection by integrating depth images to capture richer spatial and morphological details. Gate fusion module (GFM) and Transformer encoder fusion module (TFM) are designed to effectively fuse features from the two modalities. The GFM employs a dual-gated attention mechanism to perform shallow fusion and extract complementary features, while the TFM leverages a self-attention mechanism to capture global correlations and achieve deep fusion. To support the method, high-precision automatic optical inspection equipment based on a structured light phase-shift fringe technique is developed, enabling the acquisition of both depth and RGB images for constructing a comprehensive PCBA defect dataset. Experimental results demonstrate that the proposed method achieves a mean average precision of 99.89% on the dataset. Furthermore, comparative and ablation experiments are conducted to assess the individual contributions of the GFM and TFM, confirming the effectiveness and advancement of the overall approach. This method offers a valuable reference for improving surface defect detection in PCBA applications.

In conventional infrared device systems, the von Neumann architecture, which integrates sensors, processors, and memory, faces limitations due to high power consumption and high latency. Drawing inspiration from mammalian bipolar cells, this paper designs a metal-semiconductor-metal (MSM) structured infrared sensing-computing integrated device utilizing black phosphorus-indium arsenide-black phosphorus (BP-InAs-BP) as the semiconductor material. This device achieves symmetric positive/negative photoresponses and amplitude modulation across both infrared and visible light spectra through bias adjustment, enabling integrates perception and computation at the detector level. Extensive fittings of physical models facilitate the establishment of a physical simulation model for this detector, allowing microscopic-level analysis of its design, including theoretical derivation of energy band structures and micro-scale charge distributions. This significantly advances the development of BP-based infrared detector applications. In convolutional neural network tests based on this design, classification accuracy for digits 0?9 classification tasks in MNIST dataset with scale of 16 pixel×16 pixel exceeded 92%, highlighting the superior performance of this proposed infrared sensing-computing device. The simulation methodology presented in this study provides a novel design framework and theoretical analysis approach for BP material applications, offering a practical solution for neuromorphic visual perception.

This paper proposes a deep learning-based demodulation method for all-fiber Mach-Zehnder interferometer (MZI). By leveraging convolutional neural network (CNNs) to establish a nonlinear mapping model between transmission spectral curvature and applied curvature, breaking through the dynamic range limitation caused by spectral saturation in traditional demodulation methods. Two intensity-modulated fiber sensors (two-path type and in-line type) are fabricated, demonstrating original curvature demodulation ranges of 0.05747?0.10449 m-1 and 0.02612?0.49106 m?1, respectively. To achieve extended dynamic range demodulation, a Gramian angular field (GAF) encoding technique is introduced to transform one-dimensional spectral signals into two-dimensional images. The CNN regression architecture implements a neural network structure with progressively decreasing neuron counts in fully connected layers, replacing conventional classification output layers to establish nonlinear spectral-curvature mapping. Experimental results demonstrate that under constant maximum sensitivity conditions, both sensor types achieve expanded curvature measurement range of 0?1.5672 m?1, has been increased to 33 times and 3 times the original level, respectively. Validation across four network architectures (ResNet, GoogleNet, etc.) confirm the universality of this proposed method in overcoming traditional spectral limitations in fiber sensing, establishing a novel methodological framework for deep learning-enhanced fiber optic sensing technology.

As a novel optoelectronic information material, metasurfaces face a primary scientific challenge: maximizing information capacity while minimizing cost. To address this, this paper proposes a phase modulation method that combines geometric phase and propagation phase using minimalist metasurfaces composed of only two nanostructures. This design achieves holographic decoupling across three wavelength channels, enabling low-crosstalk tri-channel display performance despite limited degrees of freedom. Furthermore, we present a multi-key three-wavelength encryption scheme that integrates conventional image encryption algorithm. This approach significantly enhances information security and reliability. The designed dual-structure metasurfaces offer advantages of simple manufacturing, flexible design capability, and high information storage density, providing new solutions for image display, information encryption, anti-counterfeiting, and high-density optical storage applications.

To meet the development requirements of high-capacity fiber optic communication systems, this study employs an inverse design method to develop and implement a 0° perfect vertical grating coupler (PVGC) for multi-core fiber communication on a standard 220 nm silicon-on-insulator platform. The adjoint optimization-based inverse design approach effectively addresses multiparameter optimization challenges, demonstrating advantages of high efficiency, strong robustness, and excellent process compatibility for the optimized device. Analysis results show that the designed PVGC achieves a simulated coupling efficiency of -2.98 dB (with a 3 dB bandwidth of 50 nm) and a measured peak coupling efficiency of -3.89 dB (with a 3 dB bandwidth of 43 nm) at 1550 nm wavelength. Notably, this research has successfully developed a PVGC array for seven-core fibers, exhibiting an average peak coupling efficiency of -4.45 dB with channel uniformity errors controlled within 0.12 dB. These findings provide a high-performance and highly reliable photonic integrated solution for next-generation space-division multiplexing networks.

Optical edge detection technology has the advantages of high-speed parallel computing and high-throughput processing, and has attracted more and more attention in recent years. However, how to achieve dynamically adjustable optical edge detection is still a challenging problem. Based on this, a scheme of optical edge detection and fast switching in multiple directions is proposed. A fast tunable wave plate was fabricated based on ferroelectric liquid crystals, and a nematic liquid crystal q-plate was fabricated using photopatterning alignment technology. The experimental results show that the combination of the above two liquid crystal elements can achieve one-dimensional edge detection with high quality, and realize fast switching in two orthogonal directions as low as 35 μs by flipping the polarity of the electric field at low voltage. Based on ferroelectric liquid crystals, this scheme provides a flexible dynamic manipulation method for edge detection, which is expected to promote the application of liquid crystal elements in optical computing, optical image processing and other fields.

Conventional imaging systems based on lenses (groups) face inherent limitations such as cost, hardware size, and multifunctional applications. Lensless computational optical imaging system is a type of compact and lightweight minimalist multifunctional sensing architecture, which has been shown to be scientifically significant in a variety of fields, including biomicrography, multimodal imaging, and data security. This paper provides a concise overview of the fundamental principles, functional expansion, and intelligent reconstruction technologies underlying lensless imaging. By leveraging an"optical encoding+computational decoding"framework, this paradigm enables high-resolution, multimodal, and high-dimensional imaging, even under extreme conditions. The employment of sophisticated artificial intelligence algorithms has been demonstrated to result in a substantial enhancement of the quality of reconstructed images. This paper also discusses the potential for integrating lensless imaging and quantum intelligence technology, and anticipates its extensive applications in future endeavors, including high-security and multifunctional imaging.

Multimode fibers demonstrate significant potential in short-distance optical communications, microscopic endoscopy, and optical power transmission. However, modal dispersion causes interference and crosstalk in spatially multiplexed optical signals during transmission, affecting demodulation performance. To address this challenge, researchers have proposed various demodulation techniques, including phase conjugation and transmission matrix methods. Yet these approaches have not fully met the demands of the information age for high fidelity, interference resistance, and high-speed transmission. In multimode fiber optical information transmission systems, deep neural networks have proven effective in overcoming multiple scattering issues, enabling precise information transmission. This paper summarizes the research progress of multimode optical information transmission technologies based on deep learning, including high fidelity, high speed, anti disturbance, multi-dimensional and physical prior enhanced optical information transmission, and discusses the optimization of key parameters of the system. At the same time, the opportunities and challenges of this technology in the future are prospected.

Rapid growth in data storage and processing driven by training and inference of large-scale artificial intelligence models necessitates development of novel optical non-volatile memory materials and devices, which offer a promising solution for enhancing computational efficiency while reducing energy consumption in neural networks. Phase-change materials (PCM)-based photonic devices exhibit several advantages in big data processing with high clock frequency, large bandwidth, picosecond latency, and high energy efficiency, making it a key enabler for neuromorphic photonic computing. This review focuses on the recent advancements in optoelectronic PCM for neuromorphic computing. These PCM can be classified into several categories based on their crystallization mechanisms. We provide an in-depth discussion of their bonding mechanisms, optical properties, and performance tuning strategies. Additionally, we review the progress on PCM in photonic waveguide devices for multi-bit storage, bio-inspired synaptic behavior, neuromorphic computing, and hybrid photonic-electronic waveguide technologies. Finally, this review outlines the opportunities and challenges for the research on optoelectronic PCM.

In the intelligent era, optical perception serves as a fundamental modality for information acquisition and processing, playing a pivotal role in enabling machines to interpret the physical world. Leveraging advanced camera systems, sensors, and data processing technologies, optical systems transform rich real-world scenes into digital signals. These signals are then subjected to feature extraction via artificial intelligence algorithms to facilitate decision-making. With the proliferation of visual tasks, sensing systems have undergone continuous advancements in comprehensive performance. Among these innovations, coded optical imaging and optical pre-sensing computing, rooted in optical encoding technologies, have emerged as transformative tools in optical perception systems. These techniques not only enhance the quality of signal acquisition but also optimize system architectures, thereby boosting operational efficiency and reducing computational power consumption. This paper begins by outlining the workflow of optical perception, then delves into the revolutionary roles of optical encoding and optical computing in advancing perception technologies. It focuses on the research progress of coded optical imaging and optical pre-sensing computing, and concludes with prospects for future development opportunities and challenges in this dynamic field.

The thermo-optic effect enables precise and dynamic modulation of refractive index distributions through controllable temperature gradients in optical materials. Capitalizing on its broadband transparency, polarization independence, and non-mechanical tuning capability, this technology overcomes limitations of conventional optical modulation methods and facilitates the development of high-stability, low-power intelligent optical systems. This study systematically reviews fundamental principles, key technologies, and recent progress in thermally tunable optical devices, emphasizing the applications and achievements of two thermal modulation schemes, laser-driven and electrical-driven in tunable-focus lenses, wavefront shaping, and beam shaping. We present a comprehensive comparison of their performance characteristics and application scenarios. Additionally, this study discusses potential applications of thermally tunable optical technology in adaptive optics correction, high-resolution microscopic imaging, smart wearable devices, and precision optical fabrication, proposing theoretical frameworks and technical roadmaps for next-generation intelligent photonic devices.

The rapid development of the new generation of information technology, such as generative artificial intelligence, large models, and deep learning, has led to explosive growth in global data traffic, which puts forward higher requirements for computing power and energy consumption. Brain-inspired computing is committed to using the brain's structure, function and low-power information processing mechanism for reference to develop new information processing modes, computing models, algorithms and intelligent systems to effectively alleviate the current pressure on computing power and energy consumption. Among them, pulse neural network has many advantages, such as sparse coding, low power consumption, outstanding spatio-temporal information processing ability, and biological rationality. Optical pulse neural network further integrates the advantages of pulse neural network and photonics, such as high speed, large bandwidth, low energy consumption, and strong parallel processing ability, and has become a hot research topic. This paper reviews the work of major research teams at home and abroad in the modeling of photonic pulse neurons, device development and dynamic characteristics research, the model architecture of photonic pulse neural networks, integrated chips and other aspects, and looks forward to the challenges and future development directions.

Single-pixel imaging, a computational imaging technique utilizing wide-field illumination encoding and single-point detector sampling, offers a novel alternative to conventional imaging methods. However, the limited imaging quality and long imaging time limit the further development of single-pixel imaging in practical applications to some extent. Recent years have witnessed significant advancements in deep learning-driven single-pixel imaging, particularly in enhancing image quality and reconstruction speed. This paper elucidates the fundamental principles of deep learning and single-pixel imaging. We systematically categorize deep learning imaging methods and image-free sensing techniques in single-pixel imaging from a data mapping perspective. Additionally, we examine the advantages and limitations of both deep learning imaging methods and image-free sensing from an application standpoint. Furthermore, we comprehensively analyze the challenges facing deep learning in single-pixel imaging, explore potential solutions, and provide insights for future developments in this field.

Computational ghost imaging (CGI) achieves high-precision image reconstruction by performing second-order correlation operations between modulated optical fields and the corresponding intensity information, overcoming the limitations of traditional "point-to-point" imaging methods. This technique can decouple high-resolution object images from one-dimensional intensity signals, demonstrating high sensitivity and strong anti-interference capabilities. It holds broad application prospects in fields such as medical imaging, microscopic imaging, and LiDAR. This paper provides a detailed overview of the development and applications of traditional CGI, compressed sensing-based CGI, and deep learning-based CGI. It also analyzes the algorithms of each type of CGI and discusses the feasibility of applying large language model to CGI.

Intelligent optical computing chips have emerged as a promising solution for next-generation artificial intelligence hardware due to their high-speed broadband parallel processing capabilities, low energy-consumption, and low-latency computational characteristics. However, these chips face challenges including accumulated phase errors in complex optical paths and manufacturing process variations, necessitating high-precision, wide-bandwidth characterization and calibration technologies to achieve accurate control and practical implementation. Capitalizing on the performance advantages of photonic devices, optical vector analysis (OVA) techniques enable high-accuracy measurement, ultra-broadband characterization, and multi-dimensional analys of intelligent optical computing chips, thus serving as a crucial enabler for their applications. This paper systematically reviews the system architectures and operational principles of existing OVA technical approaches. By examining the design features and functional requirements of intelligent optical computing chips, we provide an in-depth analysis of the strengths and limitations of different technical routes. Furthermore, we discuss future development directions for OVA techniques and their application prospects in photonic device characterization and optical chip calibration.

With the explosive growth of information in modern society and the rapid development of artificial intelligence technology, intelligent computing centers are facing unprecedented challenges in terms of communication bandwidth and energy consumption. The traditional high-speed interconnection architecture has been difficult to meet the demand of the continuous growth of data traffic. In this context, silicon photonic technology continues to mature, which has the advantages of high bandwidth density and low power consumption, and is reshaping the interconnection system of modern data centers. By deeply integrating the photoelectric chips and integrating them into one, this technology greatly shortens the electrical connection distance, gradually extends the optical interconnection to the inside of the switch package, and even realizes the low delay and long-distance communication between the computing chips, which greatly expands the boundary of computing power. This paper summarizes the latest research progress of optoelectronic integrated chips for computational optical interconnection, systematically analyzes the technical challenges faced by various key devices and chips, and looks forward to the future development direction of silicon optical engine and its wide application prospect in many fields.

Facing the challenge of efficiently and accurately solving large-scale combinatorial optimization problems, spatial photonic Ising machines, by leveraging capabilities of space-division multiplexing for large-scale implementation, an intuitive and straightforward fully-connected Ising mapping mechanism, and programmable combinatorial optimization approaches, have emerged as a significant technological path for the deployment and validation of Ising machines, receiving widespread attention from both academia and industry. This paper introduces the photonic Ising mapping principle of spatial photonic Ising machines, the main technical challenges, and future research prospects. It especially focuses on three key technical issues: flexible spatial photonic Ising mapping methods, general-purpose spatial photonic Ising machine architecture, and noise-enhanced photonic Ising optimization methods.

Against the backdrop of continuously growing computing power demands and the physical limitations of Moore's Law, optical neural networks have emerged as a promising candidate for next-generation high-efficiency computing platforms due to their advantages in high parallelism, low energy consumption, and high-speed computation. Micro-nano optical neural networks, which deeply integrate micro-nano fabrication technologies with the principles of optical information processing, enable fast and energy-efficient neural network computation within extremely compact dimensions, showing great application potential. This paper first reviews the development of neural networks and the fundamental concepts of optical neural networks, and then systematically summarizes recent research advances in micro-nano optical neural networks based on waveguide propagation and free-space propagation. Finally, the key system components and architectural designs are thoroughly analyzed, and the performance characteristics of optical neural networks and electronic chips are compared across three dimensions: computational power, energy efficiency, and integration density. A systematic evaluation of the differences between cutting-edge optoelectronic hybrid chips and traditional electronic chips is also conducted. Furthermore, the challenges facing future technological development and potential breakthrough paths are proposed.

With the rapid development of artificial intelligence, traditional electronic neural networks can no longer meet the needs of large-scale data computing, so photonic neural network has emerged. With the advantages of parallel transmission of optical signals and ultra-low energy consumption, photonic neural networks can realize the synchronous processing and high-bandwidth transmission of multi-dimensional information based on optical computing, which demonstrates great potential in improving computing efficiency. Effective training methods are crucial for improving the performance of photonic neural networks, which attracts many teams to conduct research. This study reviews the work of major research teams at home and abroad in the training of photon neural networks, elaborates on the current main training methods, and focuses on analyzing training based on pulse timing dependent plasticity, gradient based training, evolutionary algorithms, hardware perception training, hardware error correction, online learning, and Python training frameworks for photon neural networks. Finally, the principles, advantages, and limitations of each method are analyzed, and prospects for the future development trends in this field are proposed.

The development of artificial intelligence has exceeded the capability boundary of the traditional von Neumann architecture computer, and the rise of neural morphology computing has broken through the computational limitations of traditional standard hardware. Optical neural network has become an ideal platform for neural morphology computing with ultra-large bandwidth, ultra-high speed, ultra-low power consumption and natural parallel computing ability. As a high-performance integrated photonic device, integrated microcavity has the nonlinear response, wavelength sensitivity and optical storage ability required to simulate biological neuron mechanism, which strongly promotes the application of wavelength division multiplexing technology in neural networks. In recent years, on-chip optical neural networks based on integrated microcavity structure have attracted extensive attention of researchers. This review systematically introduces the application of integrated microcavity devices in optical neural networks, and classifies different architectures such as optical pulse neural networks, optical convolutional neural networks and optical reserve pool computing. At the same time, the new neural network architecture based on integrated microcavity optical frequency comb is discussed, and the main challenges are analyzed.

Optical computing is a novel computational architecture based on the manipulation of photons or optical fields rather than electrons, utilizing light for information encoding, transmitting, and processing. In recent years, optical computing has leveraged independent information dimensions of photons, such as polarization, frequency, and orbital angular momentum, leading to the emergence of numerous novel architectures. Concurrently, the integration of deep-learning-driven structural design has enabled these architectures to demonstrate exceptional performance in tasks including matrix operation and image processing. This paper begins with the physical foundations of optical computing, systematically summarizes and discusses the principal architectures of free-space diffractive and on-chip integrated optical computing. Furthermore, this paper concludes by highlighting the pressing challenges confronting current developments in optical computing and provides perspectives on future trends in this field.

With the profound integration of artificial intelligence and photonics technologies, intelligent photonics is evolving into a disruptive technology that looks poised to revolutionize industries and everyday life. The development of intelligent photonics finds applications in diverse fields, including biomedicine, autonomous driving, and virtual and augmented reality. Artificial intelligence (AI) is fueling a new paradigm of photonics research, providing efficient avenues for optimizing photonics design, advancing optical systems and analyzing optical information. Enabled by the maturity of deep learning, silicon-based optoelectronics, optical materials, and quantum information, photonic computing holds great potential to address the challenges faced by Moore's law and the bottlenecks of the von Neumann architecture. Future implementations of photonic computing may meet the demands for high-performance computing in the digital infrastructure of the information era, such as those posed by 5G, big data, cloud computing, and the Internet of Things. In this study, we summarize recent advances in photonic computing, including on-chip integrated optical neural networks based on micro-ring resonators, multimode interferometers, nanobeam resonators, and subwavelength diffractive units and integration of training and computation. We also highlight the progresses in diffractive neural networks enabled by diffractive optical elements and intelligent metasurfaces, as well as the developments in photonic spiking neural networks, reservoir computing, quantum photonic computing, and large-scale optoelectronic computing chips. In terms of computational optics, we review the advances across a broad range of areas, including computational imaging, microscopy, display, fiber-optic sensing, and laser technologies.

Multimodal image fusion involves the integration of information from different sensors to obtain complementary modal features. Infrared-visible image fusion is a popular topic in multimodal tasks. However, the existing methods face challenges in effectively integrating these different modal features and generating comprehensive feature representations. To address this issue, we propose a dual-branch feature-decomposition (DBDFuse) network. A dual-branch feature extraction structure is introduced, in which the Outlook Attention Transformer (OAT) block is used to extract high-frequency local features, whereas newly designed fold-and-unfold modules in the Stoken Transformer (ST) efficiently capture low-frequency global dependencies. The ST decomposes the original global attention into a product of sparse correlation maps and low-dimensional attention to capture low-frequency global features. Experimental results demonstrate that the DBDFuse network outperforms state-of-the-art (SOTA) methods for infrared-visible image fusion. The fused images exhibit higher clarity and detail retention in visual effects, while also enhancing the complementarity between modalities. In addition, the performance of infrared and visible light fusion images in downstream tasks has been improved, with mean average accuracy of 80.98% in the M3FD object detection task and mean intersection to union ratio of 63.9% in the LLVIP semantic segmentation task.

To address the challenges of restoring details in heavily occluded areas and enhancing network generalization capabilities in occluded polarized image reconstruction tasks, the research proposes a novel occluded image reconstruction model, PolarReconGAN, based on multiscale adversarial network. The proposed model integrates with polarization array imaging technology, aims to reconstruct the polarization information of occluded targets, thereby improving image quality and detail representation. We design a multiscale feature extraction module that employs a random window slicing method to prevent information loss due to image resizing, and utilizes data augmentation to enhance model generalization. Additionally, a loss function based on discrete wavelet transform is employed to further improve the reconstruction effects of image details. The experimental results demonstrate that the proposed method achieves an average structural similarity index (SSIM) of 0.7720 and an average peak signal-to-noise ratio (PSNR) of 25.2494 dB on a multi-view occluded polarization image dataset, indicating superior performance in occluded image reconstruction.

Unsupervised representation learning is a primary method for extracting distinguishable shape information from unlabeled point cloud data. Existing approaches capture global shape features of whole point clouds but often overlook local part-level details and are computationally expensive due to their reliance on whole point clouds and numerous negative samples. Inspired by the human visual mechanism of perceiving whole objects from local shapes, this study proposes an unsupervised part-level learning network, called reconstruction contrastive part (Rc-Part). First, a dataset of 40000 part point clouds is constructed by preprocessing public whole point cloud datasets. Then, Rc-Part employs contrastive learning without negative samples to capture distinguishable semantic information among parts and uses an encoder-decoder architecture to learn part structure information. Joint training with both contrastive and reconstruction tasks is then conducted. Finally, the encoder learned from the point cloud dataset is directly applied to whole-shape classification. Experiments on the PointNet backbone achieve high classification accuracies of 90.2% and 94.0% on the ModelNet40 and ModelNet10 datasets, respectively. Notably, despite containing 10000 fewer samples than the ShapeNet dataset, the part dataset achieves superior classification performance, demonstrating its effectiveness and the feasibility for neural networks to learn global point cloud data from components.