Photonic Neural Networks Research Articles

Scaling Photonic Neural Networks: A Silicon Photonic GeMM Leveraging a Time-space Multiplexed Xbar

Scaling Photonic Neural Networks: A Silicon Photonic GeMM Leveraging a Time-space Multiplexed Xbar

10 GHz Optoelectronic Circuit With Variable Sigmoid Shaped Activation Function for Photonic Neural Networks

10 GHz Optoelectronic Circuit With Variable Sigmoid Shaped Activation Function for Photonic Neural Networks

High‐Resolution Optical Convolutional Neural Networks Using Phase‐Change Material‐Based Microring Hybrid Waveguides

In the More‐than‐Moore era, the explosive growth of data and information has driven the exploration of alternative non‐von Neumann computational paradigms. Photonic neuromorphic computing has emerged as a promising approach, offering high speed, wide bandwidth, and massive parallelism. Herein, a high‐resolution optical convolutional neural network (OCNN) is introduced using phase‐change material Ge2Sb2Te5 (GST)‐based microring hybrid waveguides. This on‐chip optical computing platform integrates GST into photonic devices, enabling versatile programming and in‐memory computing capabilities. Central to this platform is a photonic convolutional computational kernel, constructed from photonic switching cells embedded with GST on a microring resonator. This programmable photonic switch leverages the refractive index modulation during the GST phase transition to achieve up to 64 discrete levels of transmission contrast, suitable for representing matrix elements in neural network algorithms with 6‐bit resolution. Using these matrix elements, an OCNN capable of performing parallelized image edge detection and digital recognition tasks with high accuracy is demonstrated. The architecture is scalable for large‐scale photonic neural networks, offering ultrahigh computational throughput, a compact design, complementary metal‐oxide‐semiconductor‐compatible fabrication, and broad bandwidth.

Deep learning with photonic neural cellular automata.

Rapid advancements in deep learning over the past decade have fueled an insatiable demand for efficient and scalable hardware. Photonics offers a promising solution by leveraging the unique properties of light. However, conventional neural network architectures, which typically require dense programmable connections, pose several practical challenges for photonic realizations. To overcome these limitations, we propose and experimentally demonstrate Photonic Neural Cellular Automata (PNCA) for photonic deep learning with sparse connectivity. PNCA harnesses the speed and interconnectivity of photonics, as well as the self-organizing nature of cellular automata through local interactions to achieve robust, reliable, and efficient processing. We utilize linear light interference and parametric nonlinear optics for all-optical computations in a time-multiplexed photonic network to experimentally perform self-organized image classification. We demonstrate binary (two-class) classification of images using as few as 3 programmable photonic parameters, achieving high experimental accuracy with the ability to also recognize out-of-distribution data. The proposed PNCA approach can be adapted to a wide range of existing photonic hardware and provides a compelling alternative to conventional photonic neural networks by maximizing the advantages of light-based computing whilst mitigating their practical challenges. Our results showcase the potential of PNCA in advancing photonic deep learning and highlights a path for next-generation photonic computers.

Ultra-high endurance silicon photonic memory using vanadium dioxide

Silicon photonics arises as a viable solution to address the stringent resource demands of emergent technologies, such as neural networks. Within this framework, photonic memories are fundamental building blocks of photonic integrated circuits that have not yet found a standardized solution due to several trade-offs among different metrics such as energy consumption, speed, footprint, or fabrication complexity, to name a few. In particular, a photonic memory exhibiting ultra-high endurance performance (>106 cycles) has been elusive to date. Here, we report an ultra-high endurance silicon photonic volatile memory using vanadium dioxide (VO2) exhibiting a record cyclability of up to 107 cycles without degradation. Moreover, our memory features an ultra-compact footprint below 5 µm with the potential for nanosecond and picojoule programming performance. Our silicon photonic memory could find application in emerging photonic applications demanding a high number of memory updates, such as photonic neural networks with in situ training.

Proposal and detailed theoretical analysis on a photonic neural network with optically pumped Spin-VCSEL spiking neurons

A dual-layer photonic spiking neural network (PSNN) was constructed, where multiple optically pumped spin vertical-cavity surface-emitting lasers (Spin-VCSELs) were proposed as spiking neurons. Based on a detailed theoretical analysis of leaky integrate-and-fire (LIF) and refractory period characteristics of Spin-VCSEL neurons, the training and testing performance for the studied PSNN was evaluated using two standard pattern classification tasks (Iris dataset, simple digit recognition). The results showed that, by selecting appropriate parameters such as frequency detuning and number of pre-synaptic neurons, etc., higher training/testing accuracies beyond 90% can be obtained. When compared with traditional electrically pumped VCSEL, a threshold reduction of up to 50% can be achieved under nanosecond scale spin relaxation time and circular polarization optical pumping, the feasibility of realizing high accuracy (88%) pattern classification near the reduced threshold was also verified. Therefore, optically pumped Spin-VCSEL neurons can become a valuable new choice for high-performance PSNN with reduced power consumption.

Human-centred physical neuromorphics with visual brain-computer interfaces

Steady-state visual evoked potentials (SSVEPs) are widely used for brain-computer interfaces (BCIs) as they provide a stable and efficient means to connect the computer to the brain with a simple flickering light. Previous studies focused on low-density frequency division multiplexing techniques, i.e. typically employing one or two light-modulation frequencies during a single flickering light stimulation. Here we show that it is possible to encode information in SSVEPs excited by high-density frequency division multiplexing, involving hundreds of frequencies. We then demonstrate the ability to transmit entire images from the computer to the brain/EEG read-out in relatively short times. High-density frequency multiplexing also allows to implement a photonic neural network utilizing SSVEPs, that is applied to simple classification tasks and exhibits promising scalability properties by connecting multiple brains in series. Our findings open up new possibilities for the field of neural interfaces, holding potential for various applications, including assistive technologies and cognitive enhancements, to further improve human-machine interactions.

Gold saturable metasurface for building a wavelength-tunable optical spiking neuron.

Plasmonic resonant metasurfaces have found many applications in nonlinear optics, such as harmonic generation, all-optical modulation, saturable absorption, etc. A saturable absorber, as a key device for pulsing emission, plays an important role in building passively Q-switched or mode-locked fiber lasers. Recently, excitable fiber lasers have attracted much attention in the area of neuromorphic photonics. In this work, a plasmonic metasurface consisting of periodic gold nanorods resonant near 1550 nm is designed and fabricated, which exhibits saturable absorption with a modulation depth of about 2.6%. The saturable metasurface is, for the first time, utilized in an excitable erbium-doped polarization-maintaining fiber laser, acting as a crucial nonlinear term for the dynamics of the optical spiking neuron. Compared to biological neurons, the artificial optical neuron possesses shorter a refractory period, faster pulse encoding capability, and changeable firing rate as a function of cavity length (up to 20 kHz in our experiment). In addition, the optical neuron is tunable in emission wavelength within the range from 1526.3 nm to 1568.2 nm, beneficial to wavelength-division multiplexing in photonic neural networks. The trial of the nonlinear plasmonic metasurface for an excitable laser could inspire new perspectives in constructing optical neurons and extend applications of metasurfaces from conventional nonlinear optics to neuromorphic computing.

Control-free and efficient integrated photonic neural networks via hardware-aware training and pruning

Integrated photonic neural networks (PNNs) are at the forefront of AI computing, leveraging light’s unique properties, such as large bandwidth, low latency, and potentially low power consumption. Nevertheless, the integrated optical components are inherently sensitive to external disturbances, thermal interference, and various device imperfections, which detrimentally affect computing accuracy and reliability. Conventional solutions use complicated control methods to stabilize optical devices and chip, which result in high hardware complexity and are impractical for large-scale PNNs. To address this, we propose a training approach to enable control-free, accurate, and energy-efficient photonic computing without adding hardware complexity. The core idea is to train the parameters of a physical neural network towards its noise-robust and energy-efficient region. Our method is validated on different integrated PNN architectures and is applicable to solve various device imperfections in thermally tuned PNNs and PNNs based on phase change materials. A notable 4-bit improvement is achieved in micro-ring resonator-based PNNs without needing complex device control or power-hungry temperature stabilization circuits. Additionally, our approach reduces the energy consumption by tenfold. This advancement represents a significant step towards the practical, energy-efficient, and noise-resilient implementation of large-scale integrated PNNs.

Topologically-protected optical bistability based on two-dimensional photonic crystal L6 nanocavity dimer array.

We numerically investigate the optical bistability from a two-dimensional photonic crystal L6 nanocavity dimer array structure configured under the Su-Schrieffer-Heeger model. The localized electric field in the topological edge state is highly enhanced, which gives rise to strong nonlinear phenomena such as optical bistability. In comparison, a topologically trivial nanocavity is also designed and its field strength distribution and optical bistable response are also simulated. In order to test the robustness, three types of defects and interferences are introduced in both the topologically non-trivial and trivial cavities. Benefiting from the topological feature, the proposed topological cavity exhibits superior optical bistable performance with low threshold power and high switching contrast compared to that in the trivial cavity. Our work suggests what we believe to be a novel avenue toward the insertion of optical bistable devices with high robustness into future photonic integrated circuits and photonic neural networks.

Fundamentals and recent developments of free-space optical neural networks

Machine learning with artificial neural networks has recently transformed many scientific fields by introducing new data analysis and information processing techniques. Despite these advancements, efficient implementation of machine learning on conventional computers remains challenging due to speed and power constraints. Optical computing schemes have quickly emerged as the leading candidate for replacing their electronic counterparts as the backbone for artificial neural networks. Some early integrated photonic neural network (IPNN) techniques have already been fast-tracked to industrial technologies. This review article focuses on the next generation of optical neural networks (ONNs), which can perform machine learning algorithms directly in free space. We have aptly named this class of neural network model the free space optical neural network (FSONN). We systematically compare FSONNs, IPNNs, and the traditional machine learning models with regard to their fundamental principles, forward propagation model, and training process. We survey several broad classes of FSONNs and categorize them based on the technology used in their hidden layers. These technologies include 3D printed layers, dielectric and plasmonic metasurface layers, and spatial light modulators. Finally, we summarize the current state of FSONN research and provide a roadmap for its future development.

Analysis of Optical Loss and Crosstalk Noise in MZI-Based Coherent Photonic Neural Networks

Analysis of Optical Loss and Crosstalk Noise in MZI-Based Coherent Photonic Neural Networks

Ultra-fast GST-based optical neuron for the implementation of integrated photonic neural networks

A fast-integrated optical neuron is proposed based on a micro-ring resonator structure. The GST phase change material is used in the MRR structure, and neuron switching is achieved by the electrical actuation of this material. The nonlinear activation function of the neuron is extracted through three-dimensional electrical and optical simulations. The neuron is then employed in a three-layer perceptron network using the extracted activation function, and its performance is examined in the classification of handwritten digits. The overall performance of the proposed neuron is compared with the previous optical neural networks. Simulations show an ultra-fast switching time of 200fs, which is one order of magnitude faster than the best-reported value, and an accuracy of 98.9% (in the classification of MNIST handwritten digits dataset) which is among the best results, while the estimated neuron footprint is relatively small.

Inverse design of convolutional neural networks via nanophotonic kernels

We proposed an ultra-compact neural network based on inversely designed photonic kernels, which has the advantages of a small footprint, high computing speed, and low power consumption. The core layers of the neural network, composed of automatically optimized photonic multiplication kernels and addition kernels, can perform arbitrary photonic computations at the speed of the light wave, offering more advantages than their electronic counterparts. Furthermore, the convolution, pooling, and fully-connected layers can be realized by the appropriate permutation of the proposed optical kernels. We constructed an all-optical convolutional neural network with a high accuracy of 97.7% for recognizing handwritten digits. Our results could significantly promote the inverse design of photonic neural networks.

Ultralow-threshold dual-wavelength optical bistability from a perovskite hyperbolic metasurface and its application in a photonic neural network

In a hyperbolic metamaterial or metasurface (HMM or HMS), strong light-matter interaction occurs at the phase transition wavelength where the material behavior changes from metal to dielectric. As a result, the enhanced electric field can generate non-linear phenomena, such as optical bistability. In this work, we numerically investigate polarization-dependent dual-wavelength optical bistability from perovskite HMS consisting of alternating layers of MAPbBr3 perovskite and Au. Benefiting from the anisotropic property, the bistable operation wavelength for TE and TM polarizations are different. In particular, for TE polarization, strong light matter interaction resulting from the increment of incident light power not only shifts the phase transition wavelength but also enhances the transmission. Consequently, the bistable on-off contrast ratio is unprecedentedly enlarged. For TM polarization, strong light-matter interaction gives rise to an ultra-low bistable threshold. The proposed optical bistable states can serve as an activation function for optical neurons in a photonic neural network, with similar prediction accuracy across a variety of learning tasks as the classic activation functions ReLU and Sigmoid. Our work suggests a novel avenue towards the insertion of perovskite HMS into optical computing networks.

Radio‐Frequency Linear Analysis and Optimization of Silicon Photonic Neural Networks

Broadband analog signal processors utilizing silicon photonics have demonstrated a significant impact in numerous application spaces, offering unprecedented bandwidths, dynamic range, and tunability. In the past decade, microwave photonic techniques have been applied to neuromorphic processing, resulting in the development of novel photonic neural network architectures. Neuromorphic photonic systems can enable machine learning capabilities at extreme bandwidths and speeds. Herein, low‐quality factor microring resonators are implemented to demonstrate broadband optical weighting. In addition, silicon photonic neural network architectures are critically evaluated, simulated, and optimized from a radio‐frequency performance perspective. This analysis highlights the linear front‐end of the photonic neural network, the effects of linear and nonlinear loss within silicon waveguides, and the impact of electrical preamplification.

Silicon photonic neuromorphic accelerator using integrated coherent transmit-receive optical sub-assemblies

Neural networks, having achieved breakthroughs in many applications, require extensive convolutions and matrix-vector multiplication operations. To accelerate these operations, benefiting from power efficiency, low latency, large bandwidth, massive parallelism, and CMOS compatibility, silicon photonic neural networks have been proposed as a promising solution. In this study, we propose a scalable architecture based on a silicon photonic integrated circuit and optical frequency combs to offer high computing speed and power efficiency. A proof-of-concept silicon photonics neuromorphic accelerator based on integrated coherent transmit–receive optical sub-assemblies, operating over 1TOPS with only one computing cell, is experimentally demonstrated. We apply it to process fully connected and convolutional neural networks, achieving a competitive inference accuracy of up to 96.67% in handwritten digit recognition compared to its electronic counterpart. By leveraging optical frequency combs, the approach’s computing speed is possibly scalable with the square of the cell number to realize over 1 Peta-Op/s. This scalability opens possibilities for applications such as autonomous vehicles, real-time video processing, and other high-performance computing tasks.

Exploring Types of Photonic Neural Networks for Imaging and Computing-A Review.

Photonic neural networks (PNNs), utilizing light-based technologies, show immense potential in artificial intelligence (AI) and computing. Compared to traditional electronic neural networks, they offer faster processing speeds, lower energy usage, and improved parallelism. Leveraging light's properties for information processing could revolutionize diverse applications, including complex calculations and advanced machine learning (ML). Furthermore, these networks could address scalability and efficiency challenges in large-scale AI systems, potentially reshaping the future of computing and AI research. In this comprehensive review, we provide current, cutting-edge insights into diverse types of PNNs crafted for both imaging and computing purposes. Additionally, we delve into the intricate challenges they encounter during implementation, while also illuminating the promising perspectives they introduce to the field.

A photonic crystal waveguide intersection using phase change material for optical neuromorphic synapses

This paper introduces an innovative photonic crystal device that amalgamates a low cross-talk GaAs waveguide intersection with a ring employing phase change material (PCM) for application in all-optical neuromorphic synapses. The device uses a degenerate optical cavity housing Germanium–Antimony-Telluride (GST) phase-change material whose optical properties can be manipulated via a control signal. The proposed design facilitates the development of a non-volatile synapse for photonic neural networks and optical neuromorphic circuits. Numerical simulations using the finite difference time domain method (FDTD) exhibit a notably high-quality factor of 900 and a minimal cross-talk level of −60 dB at the intersection of two waveguides. To validate the FDTD results, the structure undergoes further simulations using the finite element method (FEM), confirming the accuracy of the initial calculations. The resultant structure achieves transmissions of 81 % and 13 % in the amorphous and fully crystalline states of PCM, respectively, enabling low output power in the crystalline state and high output power in the amorphous state. Moreover, modulation of the transmission coefficient between 13 % and 81 % is feasible by manipulating the crystallization coefficient of GST materials. To reduce the effect of the imaginary part of the refractive index of the GST material, the evaluation of the output transmission results for the GSST material (Ge2Sb2Se4Te1) has been performed. The results revealed that, despite having a lower imaginary part of the refractive index compared to GST, GSST exhibits lower absorption in both amorphous and crystalline states. However, the use of GSST materials shows a 27 % reduction in the figure of merit (FOM), indicating a lower effective range and fewer degrees of freedom in the different crystallization levels of the PCM ring. To investigate the effect of fabrication defects and structural sensitivity, the radius of photonic crystal rods and the rings of GaAs and GST have been changed, and the results indicate that the radius of the photonic crystal rods and the GST ring have negligible effects on the output transmission and resonant wavelength. However, the thickness of the GaAs ring shows a high sensitivity, which leads to a change in the resonance wavelength that is utilized as an effective tool for tuning the pass wavelength through the structure within the 171 nm bandgap range. The transmission loss in the amorphous state is −0.6 dB and in the fully crystalline state −7.5 dB, which is attributed to the absorption of GST material. This proposed photonic crystal synapse structure with an area of less than 30 μm2 significantly minimizes the required space compared to similar silicon photonic structures and increases the prospects for integration and scalability. The compatibility of the materials employed with optical fibre communications at the 1310 nm wavelength further bolsters the practical feasibility of this design. In summary, this paper presents a significant advancement in the domain of neuromorphic photonics by introducing a promising structure with substantial potential for neural network synapses.

Wafer-Scale Periodic Poling of Thin-Film Lithium Niobate.

Periodically poled lithium niobate on insulator (PPLNOI) offers an admirably promising platform for the advancement of nonlinear photonic integrated circuits (PICs). In this context, domain inversion engineering emerges as a key process to achieve efficient nonlinear conversion. However, periodic poling processing of thin-film lithium niobate has only been realized on the chip level, which significantly limits its applications in large-scale nonlinear photonic systems that necessitate the integration of multiple nonlinear components on a single chip with uniform performances. Here, we demonstrate a wafer-scale periodic poling technique on a 4-inch LNOI wafer with high fidelity. The reversal lengths span from 0.5 to 10.17 mm, encompassing an area of ~1 cm2 with periods ranging from 4.38 to 5.51 μm. Efficient poling was achieved with a single manipulation, benefiting from the targeted grouped electrode pads and adaptable comb line widths in our experiment. As a result, domain inversion is ultimately implemented across the entire wafer with a 100% success rate and 98% high-quality rate on average, showcasing high throughput and stability, which is fundamentally scalable and highly cost-effective in contrast to traditional size-restricted chiplet-level poling. Our study holds significant promise to dramatically promote ultra-high performance to a broad spectrum of applications, including optical communications, photonic neural networks, and quantum photonics.