arXiv subjects

Bhavin J. Shastri

Publications and source records attributed to Bhavin J. Shastri.

At least 19 recordsLinked to original sources

arXiv2026

Cryogenic thermo-optic response of low-loss phase change material for non-volatile photonic phase shifter

Phase change materials (PCMs) can enable non-volatile optical memory through the large refractive index and extinction coefficient contrast between crystalline and amorphous states. Among them, $\mathrm{Sb}_{2}\mathrm{Se}_{3}$ combines low optical attenuation at telecom wavelengths, making it promising for low-loss programmable phase shifters. Its non-volatility is particularly useful for cryogenic systems, where power dissipation and thermal load constrain scalability. However, its optical properties at cryogenic temperatures remains poorly understood. Here, we report the first cryogenic optical characterization of $\mathrm{Sb}_{2}\mathrm{Se}_{3}$ integrated on a foundry silicon photonic platform from 4K to 300K for both crystalline and amorphous states. We observe that the magnitude of the thermo-optic coefficient decreases upon cooling, whereas the optical attenuation changes only weakly for both the amorphous and crystalline phases. We report the stability of the material upon repeated thermal cycling. These measurements provide the material parameters and stability required to design low-loss, non-volatile photonic memory elements for scalable cryogenic information processing.

physics.optics↗

arXiv2026

Photonic Ising machines toward and beyond a million spins

Combinatorial optimization problems are central to many challenges in logistics, finance, engineering, and the life sciences, yet they remain among the most computationally demanding. Many of these problems can be mapped onto the Ising model, in which binary spins interact through a network of couplings, and solutions correspond to low-energy, ideally ground-state, spin configurations. Photonic Ising machines have the potential to be fast and energy-efficient heuristic solvers of optimization problems by leveraging the low latency, high bandwidth, and inherent parallelism of optics. However, current photonic implementations remain limited in scalability, connectivity, reconfigurability, and time-to-solution, preventing their use in many practical applications. In this perspective, we examine the current landscape of photonic Ising machines, discuss the challenges and limitations of existing platforms, and identify the scientific and technological advances needed to realize large-scale systems. These developments could establish photonic Ising machines as useful hardware platforms for practical optimization.

physics.optics↗

arXiv2026

Integrated electro-optic attention nonlinearities for transformers

Transformers have emerged as the dominant neural-network architecture, achieving state-of-the-art performance in language processing and computer vision. At the core of these models lies the attention mechanism, which requires a nonlinear, non-negative mapping using the Softmax function. However, although Softmax operations account for less than 1% of the total operation count, they can disproportionately bottleneck overall inference latency. Here, we use thin-film lithium niobate (TFLN) Mach-Zehnder modulators (MZMs) as analog nonlinear computational elements to drastically reduce the latency of nonlinear computations. We implement electro-optic alternatives to digital Softmax and Sigmoid, and evaluate their performance in Vision Transformers and Large Language Models. Our system maintains highly competitive accuracy, even under aggressive 4-bit input-output quantization of the analog units. We further characterize system noise at encoding speeds up to 10 GBaud and assess model robustness under various noise conditions. Our findings suggest that TFLN modulators can serve as nonlinear function units within hybrid co-packaged hardware, enabling high-speed and energy-efficient nonlinear computation.

cs.LG↗

arXiv2026

Neuromorphic Photonic Computing with an Electro-Optic Analog Memory

In neuromorphic photonic systems, device operations are typically governed by analog signals, necessitating digital-to-analog converters (DAC) and analog-to-digital converters (ADC). However, data movement between memory and these converters in conventional von Neumann architectures incur significant energy costs. We propose an analog electronic memory co-located with photonic computing units to eliminate repeated long-distance data movement. Here, we demonstrate a monolithically integrated neuromorphic photonic circuit with on-chip capacitive analog memory and evaluate its performance in machine learning for in situ training and inference using the MNIST dataset. Our analysis shows that integrating analog memory into a neuromorphic photonic architecture can achieve over 26x power savings compared to conventional SRAM-DAC architectures. Furthermore, maintaining a minimum analog memory retention-to-network-latency ratio of 100 maintains >90% inference accuracy, enabling leaky analog memories without substantial performance degradation. This approach reduces reliance on DACs, minimizes data movement, and offers a scalable pathway toward energy-efficient, high-speed neuromorphic photonic computing.

cs.ET↗

arXiv2026

Compact, Large-Scale Photonic Neurons by Modulation-and-Weight Microring Resonators

Neuromorphic photonics promises sub-nanosecond latency, ultrawide bandwidth, and high parallelism, but practical scalability is constrained by fabrication tolerances, spectral alignment, and tuning energy. Here, we present a large-scale, compact, and reconfigurable photonic neuron in which each microring performs modulation and weighting simultaneously. By exploiting both carrier and thermal tuning within a single device, this architecture reduces footprint, relaxes spectral alignment requirements to just two optical components, and yields a steep transfer response that lowers tuning energy. The proposed neuron supports multiple operating configurations, allowing its dynamical behavior to be adapted to different computational tasks. In particular, a short electrical feedback path enables recurrent operation, providing tunable short- and long-term memory for temporal processing. Using a 10-microring resonator array, we demonstrate both spatial and temporal computing, including a 3$\times$3 convolution for image processing with an error of $<$5\% and high-frequency financial time-series prediction. Each modulation-weighting element occupies 80$\times$45 \SI{}{\micro\meter^2} and consumes an average of \SI{0.186}{\milli\watt}, corresponding to a compute density of \SI{4.67}{TOPS/s/\milli\meter^2}. Excluding electronic power, the on-chip tuning efficiency reaches approximately \SI{105}{TOPs/\watt}, which is comparable to state-of-the-art implementations. These results indicate that modulation-and-weighting microring resonator banks provide a scalable building block for large-scale neuromorphic photonic systems, offering a favorable combination of compact footprint, low power consumption, and functional flexibility.

physics.optics↗

arXiv2025

Ising accelerator with a reconfigurable interferometric photonic processor

The general-purpose programmable photonic processors offer a scalable and reconfigurable solution for a wide range of RF and optical applications. Therefore, implementing photonic Ising machines using programmable processors leverages the advantages of high speed and parallelism, enabling efficient hardware acceleration for finding ground-state solutions to combinatorial optimization problems. In this work, we demonstrate a novel programmable photonic Ising solver based on a hexagonal mesh general-purpose programmable photonic platform. The integrated system allows reconfigurable matrix multiplication and computes the Hamiltonian iteratively using an annealing algorithm that facilitates spin updates and effectively searches for the ground state. As a proof of concept, we experimentally solve two benchmark optimization problems, a fundamental three-node ferromagnetic coupling problem with external bias that demonstrates nontrivial spin interactions, and a four-node Max-Cut problem with arbitrary coupling matrices. Furthermore, to establish a large-scale capability, we emulated Ising problems with sizes up to N = 50, achieving success probabilities exceeding 80\%. Additionally, we examined the impact of errors, such as phase and coupling, on the performance of the programmable photonic Ising machine. Our general-purpose photonic Ising machine paves the way for implementing large-scale, programmable architectures for solving optimization problems.

physics.optics↗

arXiv2025

Large-Scale Tree-Type Photonic Cluster State Generation with Recurrent Quantum Photonic Neural Networks

Large, multi-dimensional clusters of entangled photons are among the most powerful resources for emerging quantum technologies, as they are predicted to enable global quantum networks or universal quantum computation. Here, we propose an entirely new architecture and protocol for their generation based on recurrent quantum photonic neural networks (QPNNs) and focusing on tree-type cluster states. Unlike other approaches, QPNN-based generators are not limited by the the coherence of quantum emitters or by probabilistic multi-photon operations, enabling arbitrary scaling only limited by loss (which, unavoidably, also affects all other methods). We show that a single QPNN can learn to perform all of the many different operations needed to create a cluster state, from photon routing to entanglement generation, all with near-perfect fidelity and at loss-limited rates, even when it is created from imperfect photonic components. Although these losses ultimately place a limit on the size of the cluster states, we show that state-of-the-art photonics should already allow for clusters of 60 photons, which can grow into the 100s with modest improvements to losses. Finally, we present an analysis of a one-way quantum repeater based on these states, determining the requisite platform quality for a global quantum network and highlighting the potential of the QPNN to play a vital role in high-impact quantum technologies.

quant-ph↗

arXiv2025

Programmable 200 GOPS Hopfield-inspired photonic Ising machine

Ising machines offer a compelling approach to addressing NP-hard problems, but physical realizations that are simultaneously scalable, reconfigurable, fast, and stable remain elusive. Quantum annealers, like D-Wave's cryogenic hardware, target combinatorial optimization tasks, but quadratic scaling of qubit requirements with problem size limits their scalability on dense graphs. Here, we introduce a programmable, stable, room-temperature optoelectronic oscillator (OEO)-based Ising machine with linear scaling in spin representation. Inspired by Hopfield networks, our architecture solves fully-connected problems with up to 256 spins (65,536 couplings), and $>$41,000 spins (205,000+ couplings) if sparse. Our system leverages cascaded thin-film lithium niobate modulators, a semiconductor optical amplifier, and a digital signal processing (DSP) engine in a recurrent time-encoded loop, demonstrating potential $>$200 giga-operations per second for spin coupling and nonlinearity. This platform achieves the largest spin configuration in an OEO-based photonic Ising machine, enabled by high intrinsic speed. We experimentally demonstrate best-in-class solution quality for Max-Cut problems of arbitrary graph topologies (2,000 and 20,000 spins) among photonic Ising machines and obtain ground-state solutions for number partitioning and lattice protein folding - benchmarks previously unaddressed by photonic systems. Our system leverages inherent noise from high baud rates to escape local minima and accelerate convergence. Finally, we show that embedding DSP - traditionally used in optical communications - within optical computation enhances convergence and solution quality, opening new frontiers in scalable, ultrafast computing for optimization, neuromorphic processing, and analog AI.

physics.optics↗

arXiv2025

Fully analog end-to-end online training with real-time adaptibility on integrated photonic platform

Analog neuromorphic photonic processors are uniquely positioned to harness the ultrafast bandwidth and inherent parallelism of light, enabling scalability, on-chip integration and significant improvement in computational performance. However, major challenges remain unresolved especially in achieving real-time online training, efficient end-to-end anolog systems, and adaptive learning for dynamical environmental changes. Here, we demonstrate an on-chip photonic analog end-to-end adaptive learning system realized on a foundry-manufactured silicon photonic integrated circuit. Our platform leverages a multiplexed gradient descent algorithm to perform in-situ, on-the-fly training, while maintaining robustness in online tracking and real-time adaptation. At its core, the processor features a monolithic integration of a microring resonator weight bank array and on-chip photodetectors, enabling direct optical measurement of gradient signals. This eliminates the need for high-precision digital matrix multiplications, significantly reducing computational overhead and latency, an essential requirement for effective online training. We experimentally demonstrate real-time, end-to-end analog training for both linear and nonlinear classification tasks at gigabaud rates, achieving accuracies of over 90\% and 80\%, respectively. Our analog neuromorphic processor introduces self-learning capabilities that dynamically adjust training parameters, setting the stage for truly autonomous neuromorphic architectures capable of efficient, real-time processing in unpredictable real-world environments. As a result, we showcase adaptive online tracking of dynamically changing input datasets and achieve over 90\% accuracy, alongside robustness to external temperature fluctuations and internal thermal crosstalk.

physics.optics↗

arXiv2025

Online training and pruning of multi-wavelength photonic neural networks

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 a 96% classification accuracy, 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.

physics.optics↗

arXiv2025

Large-scale artificial intelligence with 41 million nanophotonic neurons on a metasurface

Conventional integrated circuits (ICs) struggle to meet the escalating demands of artificial intelligence (AI). This has sparked a renewed interest in an unconventional computing paradigm: neuromorphic (brain-inspired) computing. However, current neuromorphic systems face significant challenges in delivering a large number of parameters (i.e., weights) required for large-scale AI models. As a result, most neuromorphic hardware is limited to basic benchmark demonstrations, hindering its application to real-world AI challenges. Here, we present a large-scale optical neural network (ONN) for machine learning acceleration, featuring over 41 million photonic neurons. This system not only surpasses digital electronics in speed and energy efficiency but more importantly, closes the performance gap with large-scale AI models. Our ONN leverages an innovative optical metasurface device featuring numerous spatial modes. This device integrates over 41 million meta-atoms on a 10 mm$^2$ metasurface chip, enabling the processing of tens of millions of weights in a single operation. For the first time, we demonstrate that an ONN, utilizing a single-layer metasurface, can match the performance of deep and large-scale deep learning models, such as ResNet and Vision Transformer, across various benchmark tasks. Additionally, we show that our system can deliver high-performance solutions to real-world AI challenges through its unprecedented scale, such as accelerating the analysis of multi-gigapixel whole slide images (WSIs) for cancer detection by processing the million-pixel sub-image in a single shot. Our system reduces computing time and energy consumption by over 1,000 times compared to state-of-the-art graphic processing units (GPUs). This work presents a large-scale, low-power, and high-performance neuromorphic computing system, paving the way for future disruptive AI technologies.

physics.optics↗

arXiv2025

Photoluminescent colour centres on a mainstream silicon photonic foundry platform

The fabrication of silicon photonic components in commercial CMOS-compatible foundries has revolutionized the impact of silicon photonics on advancing communication, quantum computing and artificial intelligence, due to their benefits of mass production, high throughput, low cost, and high performance. The indirect bandgap of silicon introduces a fundamental challenge; thus, the mainstream silicon-on-insulator (SOI) platform does not have efficient light sources. Recently, luminescent colour centres in SOI have emerged as one promising approach for developing efficient on-chip classical and quantum light sources, although past work has relied on custom fabrication that is not foundry-compatible. In this work, we demonstrate W-centre photoluminescence on a mainstream silicon photonics platform through development of a straightforward back end-of-line (BEOL) treatment. At an optimal implant energy of 7~MeV, we observed W-centre photoluminescence with a brightness comparable to prior in-house processes. We performed a series of experiments on Circular Bragg Grating (CBG) devices with varying pitches, duty cycles, and implant energies to confirm the PL emission from the encapsulated SOI device layer rather than the handle wafer. Our novel approach in fabricating silicon colour centres in commercial silicon photonic foundry processes opens up new opportunities for integrating classical and quantum light sources directly onto silicon photonic circuits, unlocking opportunities for large-scale integration of advanced photonic architectures on chip.

physics.optics↗

arXiv2025

Roadmap on Neuromorphic Photonics

This roadmap consolidates recent advances while exploring emerging applications, reflecting the remarkable diversity of hardware platforms, neuromorphic concepts, and implementation philosophies reported in the field. It emphasizes the critical role of cross-disciplinary collaboration in this rapidly evolving field.

cs.ET↗

arXiv2024

Physics to System-level Modeling of Silicon-organic-hybrid Nanophotonic Devices

The continuous growth in data volume has sparked interest in silicon-organic-hybrid nanophotonic devices integrated into silicon photonic integrated circuits (PICs). SOH devices offer improved speed and energy efficiency compared to silicon photonics devices. However, a comprehensive and accurate modeling methodology of SOH devices is lacking. While some preliminary modeling approaches for SOH devices exist, their reliance on theoretical and numerical methodologies, along with a lack of compatibility with electronic design automation (EDA), hinders their seamless and rapid integration with silicon PICs. Here, we develop a phenomenological, building-block-based SOH PICs simulation methodology spanning from the physics to the system level, offering high accuracy, comprehensiveness, and EDA-style compatibility. Our model is also readily integrable and scalable, lending itself to the design of large-scale silicon PICs. Our proposed modeling methodology is agnostic and compatible with any photonics-electronics co-simulation software. We validate this methodology by comparing experimentally demonstrated SOH microring modulators and Mach Zehnder modulators with those obtained through simulation, demonstrating its ability to model various modulator topologies. We also show our methodology's ease and speed in modeling large-scale systems. As an illustrative example, we use our methodology to design and study a 3-channel SOH MRM-based wavelength-division (de)multiplexer, a widely used component in various applications, including neuromorphic computing, data center interconnects, communications, and sensing. Our modeling approach is compatible with other materials exhibiting the Pockels and Kerr effects. To our knowledge, this represents the first comprehensive physics-to-system-level EDA-compatible simulation methodology for SOH modulators.

physics.optics↗

arXiv2024

Fully integrated hybrid multimode-multiwavelength photonic processor with picosecond latency

High-speed signal processing is essential for maximizing data throughput in emerging communication applications, like multiple-input multiple-output (MIMO) systems and radio-frequency (RF) interference cancellation. However, as these technologies scale, they increase hardware complexity, computing power demands, and create significant digital signal processing (DSP) challenges. While transistor miniaturization has improved digital electronic processors, they still face physical bottlenecks, limiting computational throughput and increasing DSP latency. Photonic processors present a promising alternative, offering large bandwidth, low loss, parallel processing, and low latency. Yet, scalability in photonic processors remains limited by system integration, device size, and on-chip multiplexing challenges. Here, we introduce a scalable on-chip hybrid multiplexed photonic processor, combining mode-division multiplexing (MDM) and wavelength-division multiplexing (WDM). This marks the first implementation of a monolithically integrated MDM-WDM-compatible processor, featuring mode multiplexers, multimode microring resonators, and multimode balanced photodetectors. Furthermore, we demonstrate real-time unscrambling of 5 Gb/s non-return-to-zero optical MIMO signals and RF phase-shift keying signal unjamming. Our system's 30 ps processing latency makes it ideal for real-time MIMO and RF applications. Our analysis reveals that hybrid MDM-WDM multiplexing improves the number of operations per second by 4.1 times over spatially multiplexed WDM configurations, positioning it as a strong candidate for next-generation large-scale photonic processors.

physics.optics↗

arXiv2024

120 GOPS Photonic Tensor Core in Thin-film Lithium Niobate for Inference and in-situ Training

Photonics offers a transformative approach to artificial intelligence (AI) and neuromorphic computing by enabling low-latency, high-speed, and energy-efficient computations. However, conventional photonic tensor cores face significant challenges in constructing large-scale photonic neuromorphic networks. Here, we propose a fully integrated photonic tensor core, consisting of only two thin-film lithium niobate (TFLN) modulators, a III-V laser, and a charge-integration photoreceiver. Despite its simple architecture, it is capable of implementing an entire layer of a neural network with a computational speed of 120 GOPS, while also allowing flexible adjustment of the number of inputs (fan-in) and outputs (fan-out). Our tensor core supports rapid in-situ training with a weight update speed of 60 GHz. Furthermore, it successfully classifies (supervised learning) and clusters (unsupervised learning) 112 * 112-pixel images through in-situ training. To enable in-situ training for clustering AI tasks, we offer a solution for performing multiplications between two negative numbers.

physics.optics↗

arXiv2024

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 on light's unique properties, such as large bandwidth, low latency, and potentially low power consumption. Nevertheless, the integrated optical components within PNNs are inherently sensitive to external disturbances and thermal interference, which can detrimentally affect computing accuracy and reliability. Current solutions often use complicated control methods, resulting in high hardware complexity impractical for large-scale PNNs. In response, we propose a novel hardware-aware training and pruning approach. The core idea is to train the parameters of a physical neural network towards its noise-robust and energy-efficient region. This innovation enables control-free and energy-efficient photonic computing. Our method is validated across diverse integrated PNN architectures. Through experimental validation, our approach significantly enhances the computing precision of MRR-based PNN, achieving a notable 4-bit improvement without the need for complex device control mechanisms or energy-intensive temperature stabilization circuits. Specifically, it improves the accuracy of experimental handwritten digit classification from 67.0% to 95.0%, nearing theoretical limits and achieved without a thermoelectric controller. Additionally, this approach reduces the energy by tenfold. We further extend the validation to various architectures, such as PCM-based PNN, demonstrating the broad applicability of our approach across different platforms. This advancement represents a significant step towards the practical, energy-efficient, and noise-resilient implementation of large-scale integrated PNNs.

physics.optics↗

arXiv2024

Nonlinear dynamics in neuromorphic photonic networks: physical simulation in Verilog-A

Advances in silicon photonics technology have enabled the field of neuromorphic photonics, where analog neuron-like processing elements are implemented in silicon photonics technology. Accurate and scalable simulation tools for photonic integrated circuits are critical for designing neuromorphic photonic circuits. This is especially important when designing networks with recurrent connections, where the dynamics of the system may give rise to unstable and oscillatory solutions which need to be accurately modelled. These tools must simultaneously simulate the analog electronics and the multi-channel (wavelength-division-multiplexed) photonics contained in a photonic neuron to accurately predict on-chip behaviour. In this paper, we utilize a Verilog-A model of the photonic neural network to investigate the dynamics of recurrent integrated circuits. We begin by reviewing the theory of continuous-time recurrent neural networks as dynamical systems and the relation of these dynamics to important physical features of photonic neurons such as cascadability. We then present the neural dynamics of systems of one and two neurons in the simulated Verilog-A circuit, which are compared to the expected dynamics of the abstract CTRNN model. Due to the presence of parasitic circuit elements in the Verilog-A simulation, it is seen that there is a topological equivalence, but not an exact isomorphism, between the theoretical model and the simulated model. The implications of these discrepancies for the design of neuromorphic photonic circuits are discussed. Our findings pave the way for the practical implementation of large-scale silicon photonic recurrent neural networks.

cs.ET↗