Monolithic Amplitude‐Phase Modulator for Scalable Optical Convolution

In this paper, we present AnalogVNN, a simulation framework built on PyTorch that can simulate the effects of optoelectronic noise, limited precision, and signal normalization present in photonic neural network accelerators. We use this framework to train and optimize linear and convolutional neural networks with up to nine layers and ∼1.7 × 106 parameters, while gaining insights into how normalization, activation function, reduced precision, and noise influence accuracy in analog photonic neural networks. By following the same layer structure design present in PyTorch, the AnalogVNN framework allows users to convert most digital neural network models to their analog counterparts with just a few lines of code, taking full advantage of the open-source optimization, deep learning, and GPU acceleration libraries available through PyTorch.

CMOS-compatible photonic integrated circuits (PICs) are emerging as a promising platform in artificial intelligence (AI) computing. Owing to the compact footprint of microring resonators (MRRs) and the enhanced interconnect efficiency enabled by wavelength division multiplexing (WDM), MRR-based photonic neural networks (PNNs) are particularly promising for large-scale integration. However, the scalability and energy efficiency of such systems are fundamentally limited by the MRR resonance wavelength variations induced by fabrication process variations (FPVs) and environmental fluctuations. Existing solutions use post-fabrication approaches or thermo-optic tuning, incurring high control power and additional process complexity. In this work, we introduce an online training and pruning method that addresses this challenge, adapting to FPV-induced and thermally induced shifts in MRR resonance wavelength. By incorporating a power-aware pruning term into the conventional loss function, our approach simultaneously optimizes the PNN accuracy and the total power consumption for MRR tuning. In proof-of-concept on-chip experiments on the Iris dataset, our system PNNs can adaptively train to maintain above 90 % classification accuracy in a wide temperature range of 26–40 °C while achieving a 44.7 % reduction in tuning power via pruning. Additionally, our approach reduces the power consumption by orders-of-magnitude on larger datasets. By addressing chip-to-chip variation and minimizing power requirements, our approach significantly improves the scalability and energy efficiency of MRR-based integrated analog photonic processors, paving the way for large-scale PICs to enable versatile applications including neural networks, photonic switching, LiDAR, and radio-frequency beamforming.

Reconfigurable silicon photonic devices are widely used in numerous emerging fields such as optical interconnects, photonic neural networks, quantum computing, and microwave photonics. Currently, phase change materials (PCMs) have been extensively investigated as promising candidates for building switching units due to their strong refractive index modulation. Here, nonvolatile multilevel switching of silicon photonic devices with Ge2Sb2Te5 (GST) is demonstrated with In2O3 transparent microheaters that are compatible with diverse material platforms. With GST integrated on the silicon photonic waveguides and Mach‐Zehnder interferometers (MZIs), repeatable and reversible multilevel modulation of GST is achieved by electro‐thermally induced phase transitions. Particularly, the segmented switching unit of In2O3 and GST is proposed and demonstrated to be capable of producing about one order of magnitude larger temperature gradient than that of the nonsegmented unit, resulting in up to 64 distinguishable switching levels of 6‐bit precision, and fine‐tuning of the switching voltage pulses is promising to push the precision even further, to 7‐bit, or 128 distinguishable switching levels. The capability of precise multilevel phase‐change modulation is crucial to further facilitate the development of nonvolatile reconfigurable switches and variable attenuation devices as building blocks in large‐scale programmable optoelectronic systems.

Digital signal processing (DSP) has been widely applied in optical communication systems to mitigate various signal distortions and has become one of the key technologies that have sustained data traffic growth over the past decade. However, the strict energy budget of application-specific integrated circuit-based DSP chips has prevented the deployment of some powerful but computationally costly DSP algorithms in real applications. As a result, fiber nonlinearity-induced signal distortions impede fiber communication systems, especially in wavelength-division multiplexed (WDM) transmission systems. To solve these challenges in DSP, there has been a surge of interest in implementing neural networks-based signal processing using photonics hardware (i.e., photonic neural networks). Photonic neural networks promise to break performance limitations in electronics and gain advantages in bandwidth, latency, and power consumption in solving intellectual tasks that are unreachable by conventional digital electronic platforms. This work proposes a photonic recurrent neural network (RNN) capable of simultaneously resolving dispersion and both intra and inter-channel fiber nonlinearities in multiple WDM channels in the photonic domain, for the first time to our best knowledge. Furthermore, our photonic RNN can directly process optical WDM signals in the photonic domain, avoiding prohibitive energy consumption and speed overhead in analog to digital converters (ADCs). Our proposed photonic RNN is fully compatible with mature silicon photonic fabrications. We demonstrate in simulation that our photonic RNN can process multiple WDM channels simultaneously and achieve a reduced bit error rate compared to typical DSP algorithms for all WDM channels in a pulse-amplitude modulation 4-level (PAM4) transmission system, thanks to its unique capability to address inter-channel fiber nonlinearities. In addition to signal quality performance, the proposed system also promises to significantly reduce the power consumption and the latency compared to the state-of-the-art DSP chips, according to our power and latency analysis.

Photonic neural networks (PNN) are a promising alternative to electronic GPUs to perform machine-learning tasks. The PNNs value proposition originates from i) near-zero energy consumption for vector matrix multiplication once trained, ii) 10-100 ps short interconnect delays, iii) weak required optical nonlinearity to be provided via fJ/bit efficient emerging electrooptic devices. Furthermore, photonic integrated circuits (PIC) offer high data bandwidth at low latency, with competitive footprints and synergies to microelectronics architectures such as foundry access. This talk discusses recent advances in photonic neuromorphic networks and provides a vision for photonic information processors. Details include, 1) a comparison of compute performance technologies with respect to compute efficiency (i.e. MAC/J) and compute speed (i.e. MAC/s), 2) a discussion of photonic neurons, i.e. perceptrons, 3) architectural network implementations, 4) a broadcast-and-weight protocol, 5) nonlinear activation functions provided via electro-optic modulation, and 6) experimental demonstrations of early-stage prototypes. The talk will open up answering why neural networks are of interest, and concludes with an application regime of PNN processors which reside in deep-learning, nonlinear optimization, and real-time processing. Photonic neural networks (PNN) are a promising alternative to electronic GPUs to perform machine-learning tasks. The PNNs value proposition originates from i) near-zero energy consumption for vector matrix multiplication once trained, ii) 10-100 ps short interconnect delays, iii) weak required optical nonlinearity to be provided via fJ/bit efficient emerging electrooptic devices. Furthermore, photonic integrated circuits (PIC) offer high data bandwidth at low latency, with competitive footprints and synergies to microelectronics architectures such as foundry access. This talk discusses recent advances in photonic neuromorphic networks and provides a vision for photonic information processors. Details include, 1) a comparison of compute performance technologies with respect to compute efficiency (i.e. MAC/J) and compute speed (i.e. MAC/s), 2) a discussion of photonic neurons, i.e. perceptrons, 3) architectural network implementations, 4) a broadcast-and-weight protocol, 5) nonlinear activation functions provided via electro-optic modulation, and 6) experimental demonstrations of early-stage prototypes. The talk will open up answering why neural networks are of interest, and concludes with an application regime of PNN processors which reside in deep-learning, nonlinear optimization, and real-time processing.

This paper reviews the latest research progress and key technologies in accelerating edge computing with optical neural networks, set against the backdrop of autonomous driving. It first outlines the context and research significance of autonomous driving, highlighting the technical limitations faced by traditional electronic computing. Subsequently, it provides a detailed analysis of the latest research achievements in optical neural networks and accelerated computing both domestically and internationally. This includes innovative applications of optical flow estimation methods (FocusFlow, CSFlow), introductions to methods for edge computing and autonomous driving collaboration (OVEAP, DVEAP), and innovative approaches to underlying computations (optical convolution, Monet channels). Through comparative analysis, it reveals current re-search hotspots and trends, analyzes existing short-comings and potential solutions, and explores possible future development directions. This paper aims to systematically explore how optical computing technology can overcome the limitations of traditional electronic computing to build a new generation of high-performance, low-power intelligent computing systems.

With the increasing computational demands of artificial intelligence, photonic neural networks have emerged as a promising alternative to electronic accelerators, offering high speed, low latency, and energy efficiency. This paper presents a low-loss crossing waveguide-assisted silicon-based microdisk resonator (MDR) crossbar array integrated with phase change material (PCM) Ge2Sb2Se4Te1 (GSST) for optical neural network applications. The designed multimode interference (MMI) crossing waveguide features a compact footprint of 1 µm2 with a 70-degree offset, reducing the device area by a factor of 2.03 while maintaining low insertion loss (IL) and crosstalk (CT). By integrating GSST on the microdisks, we achieve nonvolatile and programmable weight control via crystallization degree modulation. A 3 × 3 MDR array is demonstrated to perform photonic matrix-vector multiplication with a computational efficiency of 1.2 × 1012 MAC/s/mm2. The scalability of the system has been further validated through image convolution tasks and fall detection tasks. This work provides a scalable and efficient architecture for on-chip photonic computing.

Neural networks (NNs) have been widely applied in computer vision, natural language processing, and beyond. As the miniaturization of semiconductor transistors nears its limits, thereby bringing Moore’s law to an end, photonic neural networks (PNNs), which leverage photons rather than electrons as information carriers, have emerged as promising hardware accelerators for NNs. PNNs offer ultrafast processing speed, ultra-low energy consumption, and extremely high throughput. Moreover, advances in photonic integrated circuits provide compact and reliable hardware platforms for constructing PNNs. Here, we provide an overview of recent advances in integrated PNN components, including linear and nonlinear operators based on diverse optical elements, as well as devices that enable optical fan-in and fan-out. We further summarize the development in on-chip PNN implementations and discuss the future challenges associated with their monolithic integration.

Microdisk modulator-assisted optical nonlinear activation functions for photonic neural networks

Miniaturization and performance enhancement of photonic devices remain key challenges in the advancement of optical information processing technologies and photonic neural networks. In this study, an inverted triangular resonator (ITR) based on a metal–insulator–metal (MIM) plasmonic architecture and phase-change material Ge2Sb2Te5 (GST) is introduced. By integrating a GST element with a volume of only 78,000 nm³ into the active region of the ITR, strong electromagnetic field confinement and optimized phase transition control are achieved. This design results in an unprecedentedly small footprint (0.031 μm²) and an outstanding optical contrast of 76% at the telecommunication wavelength of 1598 nm. The switching energy of this structure, estimated by numerical simulations, is about 7 pJ, placing it among the most energy-efficient photonic devices. The unique features of this structure—including the extremely small volume of active material, low energy consumption, high thermal stability, and excellent reliability—make it an ideal candidate for implementing tunable optical synapses in neuromorphic systems. Its fast nonlinear response and stable operation enable precise simulation of synaptic weighting and the implementation of learning algorithms in photonic architectures. These results demonstrate that the intelligent combination of plasmonic technology and phase-change materials in the ITR configuration can open new horizons in the design of ultra-compact, low-power photonic devices for next-generation neural networks and optical information processing.

Photonic neural networks have emerged as a promising solution to overcome limitations of traditional hardware for neuromorphic computations, offering advantages in bandwidth, latency, and power efficiency. However, their performance is constrained by the limited precision of analog photonic computing, which is affected by inherent noise sources such as thermal and shot noise, and distortions. These effects degrade the photonic neural network performance, reducing the bit resolution achievable in photonic hardware typically to 2–4 bits. Traditional quantization strategies fail to account for these noise contributions, resulting in a substantial accuracy loss during inference. This paper introduces an adaptive quantization method called Adaptive-Quantization Photonic-Aware Neural Network (AQ-PANN) to address the challenges posed by different noise sources in analog photonic hardware. The proposed method uses a learnable step size quantization scheme to achieve high accuracy and stability under varying noise levels, introducing a scheme that unifies quantization step adaptation with noise injection exactly where photonic distortions arise. This design incurs only a minor training-time overhead, as it involves learning a small number of per-layer quantization step sizes and does not affect inference. Experimental evaluations on three commonly used test datasets (MNIST, SVHN, and CIFAR-10) with different bit resolutions show the robustness of AQ-PANN. On MNIST, an accuracy drop of only 2% was observed from low to high noise levels in a 4-bit configuration, while traditional DoReFa quantization suffered a 29% drop. For the SVHN dataset, AQ-PANN obtained a mean accuracy of 92% under high noise with 4-bit quantization, outperforming DoReFa by over 45%. On CIFAR-10, AQ-PANN maintained close to 60% accuracy under high noise in the 4-bit configuration, whereas DoReFa and PACT both collapsed below 40%. These results highlight the effectiveness of AQ-PANN in sustaining model performance across different noise intensities, enabling practical photonic neural network deployment.

Convolutional neural network (CNN) and recurrent neural network (RNN) are two popular architectures used in text classification. Traditional methods to combine the strengths of the two networks rely on streamlining them or concatenating features extracted from them. In this paper, we propose a novel method to keep the strengths of the two networks to a great extent. In the proposed model, a convolutional neural network is applied to learn a 2D weight matrix where each row reflects the importance of each word from different aspects. Meanwhile, we use a bidirectional RNN to process each word and employ a neural tensor layer that fuses forward and backward hidden states to get word representations. In the end, the weight matrix and word representations are combined to obtain the representation in a 2D matrix form for the text. We carry out experiments on a number of datasets for text classification. The experimental results confirm the effectiveness of the proposed method.

Tunable optical micro-ring filters play significant roles in optical communications, microwave photonics, and photonic neural networks. Typical micro-ring filters are based on either a thermo-optic (TO) effect with microsecond timescales or an electro-optic (EO) effect with a limited tuning range. Here, we report a continuously tunable lithium niobate on insulator (LNOI) Vernier cascaded micro-ring filter with wire-bonded packaging integrated with both TO and EO tuning electrodes, featuring a 40-nm free spectral range (FSR), 2.3 GHz EO bandwidth, and a high sidelobe suppression ratio of 21.7 dB, simultaneously. Our high-performance optical micro-ring filter could become an important element in future LNOI photonic circuits with applications in high-capacity wavelength-division multiplexing (WDM) systems, broadband microwave photonics, fast-tunable external-cavity lasers, and high-speed photonic neural networks.

Abstract: Photonic Neural Networks (PNNs) represent a transformative leap in artificial intelligence, offering ultra-fast processing speeds and significant energy efficiency compared to traditional electronic models. This research explores the architecture, implementation, and performance evaluation of PNNs as a paradigm shift in optical computing and machine learning. Using experimental simulations and predictive regression models, this study evaluates the potential of PNNs to outperform electronic neural networks in processing speed, power consumption, and scalability. Key findings indicate a 4–6× improvement in processing throughput and a 70% reduction in energy consumption under standardized machine learning workloads. These insights provide a new direction for AI system development, merging photonics and neuromorphic computing. Keywords: Photonic Neural Networks, Optical Computing, Machine Learning, Energy Efficiency, AI Acceleration, Neuromorphic Computing, Regression Analysis, Predictive Modeling, Optical Hardware, High-Speed Computing

paper, we focus on exploring the relationship between weave patterns and their mechanical properties in woven fiber composites through Machine Learning. Specifically, we explore the interactions between woven architectures and in-plane stiffness properties through Deep Convolutional Neural Network (DCNN) and Generative Adversarial Network (GAN). Our research is important for exploring how woven composite’s pattern is related to its mechanical properties and accelerating woven composite design as well as optimization. We focus on two tasks: (1) Stiffness prediction: Predicting in-plane stiffness properties for given weave patterns. Our DCNN extracts high-level features through several convolutional and fully connected layers to determine the final predictions. (2) Weave pattern prediction: Predicting weave patterns for target stiffness properties, which can be treated as the reverse task of the first one. Due to many-to-one mapping between weave patterns and the composite properties, we utilize a Decoder Neural Network as our baseline model and compare its performance with GAN and Genetic Algorithm. We represent the weave patterns as 2D checkerboard models and use finite element analysis (FEA) to determine in-plane stiffness properties, which serve as input data for our ML framework. We show that: (1) for stiffness prediction, DCNN can predict stiffness values for a given weave pattern with relatively high accuracy (above 93%); (2) for weave pattern prediction, the GAN model gives the best prediction accuracy (above 92%) while Decoder Neural Network has the best time efficiency. HAOTIAN FENG