physics.opticsApr 12, 2026

Harnessing Photonics for Machine Intelligence

Authors: Hanqing Zhu, Shupeng Ning, Hongjian Zhou, Ziang Yin, Ray T. Chen, Jiaqi Gu, David Z. Pan

Organizations: The University of Texas at Austin · Arizona State University

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

The exponential growth of machine-intelligence workloads is colliding with the power, memory, and interconnect limits of the post-Moore era, motivating compute substrates that scale beyond transistor density alone. Integrated photonics is emerging as a candidate for artificial intelligence (AI) acceleration by exploiting optical bandwidth and parallelism to reshape data movement and computation. This review reframes photonic computing from a circuits-and-systems perspective, moving beyond building-block progress toward cross-layer system analysis and full-stack design automation. We synthesize recent advances through a bottleneck-driven taxonomy that delineates the operating regimes and scaling trends where photonics can deliver end-to-end sustained benefits. A central theme is cross-layer co-design and workload-adaptive programmability to sustain high efficiency and versatility across evolving application domains at scale. We further argue that Electronic-Photonic Design Automation (EPDA) will be pivotal, enabling closed-loop co-optimization across simulation, inverse design, system modeling, and physical implementation. By charting a roadmap from laboratory prototypes to scalable, reproducible electronic-photonic ecosystems, this review aims to guide the CAS community toward an automated, system-centric era of photonic machine intelligence.

Figures & tables

Fig. 1 : Physical advantages of photonic computing: ultra-low latency via RC-free propagation, high bandwidth density via multiplexing, and superior energy efficiency.
Fig. 2 : System-level energy efficiency vs. compute density comparison between silicon-photonic accelerators and digital GPUs. The axes report effective energy efficiency (TOPS/W) and compute density (TOPS/mm 2 ) under reported or derived system-level assumptions. GPU points correspond to NVIDIA V100 [ 45 ] , A100 [ 46 ] , H100 SXM [ 47 ] , and B200 [ 48 ] , using vendor-reported peak INT8 performance and die area. Photonic points are compiled from prior silicon-photonic accelerator works [ 23 , 28 , 29 , 32 , 42 , 49 , 50 , 51 , 52 , 53 , 54 , 55 , 56 , 57 , 58 , 59 , 60 , 61 , 62 , 63 , 64 , 65 , 66 , 67 , 68 , 69 ] and are annotated to reflect which scaling dimensions each design emphasizes (area efficiency, energy efficiency, robustness at scale, and dynamic-workload capability, such as attention).
Fig. 3 : SimPhony [ 33 ] cross-layer modeling framework for heterogeneous electronic-photonic AI systems. A modular PTC can be instantiated from different architecture families under a unified system interface. The framework maps NN operators into GEMM workloads, generates optics-specific dataflow, and evaluates end-to-end metrics using analyzers for memory traffic, area/floorplan, power/energy breakdown, and optical link budget.
Fig. 4 : (a) Energy efficiency and compute density comparison between photonic and digital electronic hardware. (b) End-to-end energy breakdown comparison across 3 representative PTC families under dynamic attention and static linear workloads with matched computations. All results are simulated using SimPhony [ 33 ] . Arch Setting : Total of 8 8 ×\times 8 PTCs with 12 wavelengths, 8-bit precision, and a 5 GHz clock rate. Workload : Attention uses sequence length S=1024S=1024 and embedding dimension D=512D=512 ; Linear layer projects hidden dimension from 512 to 1024 with a batch of 4. MZI mesh is incompatible with attention, thus omitted.
Fig. 5 : Impact of parameter scaling on energy breakdown and power efficiency across representative designs on a single tensor core . We evaluate three canonical photonic tensor core designs, the dynamic time-multiplexing crossbar [ 32 ] , the Clements MZI mesh [ 28 ] , and the weight bank [ 49 ] , using simplified and standardized models. Energy breakdowns (left panels) and energy efficiency in TOPS/W (right panels) are reported as functions of (a) core size, (b) number of wavelengths, (c) bit precision, and (d) operating frequency. The baseline configuration is one 8 ×\times 8 core, 12 wavelengths, 8-bit precision, and 5 GHz operation. All results are obtained using SimPhony [ 33 ] to ensure a fair and consistent comparison across architectures. All scaling explorations are performed using a 4-batch linear projection ( 512 ⁣→ ⁣1024512\!\rightarrow\!1024 ) to ensure all PTC families can be evaluated.
DomainApproachModelInputs | ConditionPhysics PriorDevices/ScalesSpeedupErrorComment
Frequency -domainOne-shotMaxwellNet [ 96 ]ϵr ∣ λ,Ω,J,\epsilon_{r}~|~\lambda,\Omega,J, PMLMaxwell LossLens/ <10λ<10\lambda300-600 ×\times over COMSOL∼\sim 0.01 N-L2 Error 
WaveY-Net [ 97 ] (U-Net)ϵr ∣ λ,Ω,J,\epsilon_{r}~|~\lambda,\Omega,J, PMLMaxwell LossGrating/ <10λ<10\lambda700 ×\times over Direct Solver3e-2 MAEFast Limited scalability
NeurOLight [ 98 ] (FNO)ϵr,λ,Ω,J ∣\epsilon_{r},\lambda,\Omega,J~| PMLWave PriorTunable/etched MMI/ ∼10λ\sim 10\lambda>100 ×\times over Direct Solver0.12 N-MAE
PACE [ 99 ] (FNO)ϵr,λ,Ω,J ∣\epsilon_{r},\lambda,\Omega,J~| PMLWave PriorTunable/etched MMI, Metaline/ ∼10λ\sim 10\lambda150-500 ×\times over Direct Solver0.03-0.1 N-MAE
IterativeFNO+F-GMRES [ 100 ]ϵr,λ,Ω,J,∣Ae−b∣,\epsilon_{r},\lambda,\Omega,J,|Ae-b|, PMLMaxwell ResidualWDM, Coupler, Metalens/ ∼100λ\sim 100\lambda10 ×\times over iterative GMRES1e-3 L1 ErrorScalable Limited speedup
Time -domainAutoregressivePIC 2 OSim [ 101 ] (CNN)ϵr,λ,Ω,J\epsilon_{r},\lambda,\Omega,J | PMLCausality, ModelMRR, MMI, Metaline300-600 ×\times over MEEP3e-2 N-L2NormBroadband Error accumulates
TABLE I : Categorized AI-assisted photonic device full-field simulation approaches with key feature comparison. N- means normalized. MAE is the mean absolute error.
Fig. 6 : Photonic component inverse design flows. (1) Optimization-based, (2) Generative, and (3) Hybrid methods.
DeviceGeometryInvDes methodDoFDesign regionFab-aware# Sims
Optical amplifier † [ 111 ]StructuralPSO + NN7N/A✗∼\sim 10,000
Wavelength router [ 112 ]StructuralPSO + NN5N/APost-hoc∼\sim 1,000
Phase shifter [ 113 ]StructuralPSO3Lenght ∼3 μm\sim 3~\mu\mathrm{m}✗∼\sim 1,000
Few-mode fiber [ 114 ]StructuralInversed NN5N/APost-hoc∼\sim 10,000
Microring resonator [ 115 ]BoundaryBO∼\sim 10Radius ∼3 μm\sim 3~\mu\mathrm{m}Post-hoc∼\sim 100
Mode splitter [ 116 ]BoundaryAGO∼\sim 20014×2.5 μm214\times 2.5~\mu\mathrm{m}^{2}Post-hoc∼\sim 400
TABLE II : Representative inverse-designed photonic devices, comparing geometry parameterization schemes, optimization methods, degrees of freedom (DoF), design-region sizes, fabrication-awareness strategies, and the approximate number of simulations required (# Sims). Devices marked with † are active components. Abbreviations: PSO = particle swarm optimization; BO = Bayesian optimization; GA = genetic algorithm; DRL = deep reinforcement learning; DBS = direct-binary search; AGO = adjoint gradient optimization.
Fig. 7 : Inverse-designed photonic components and circuit modules can achieve similar functionalities with orders-of-magnitude smaller spatial footprint compared to manual counterparts. (a) Manual [ 131 ] and inverse-designed four-mode mode-division multiplexer [ 132 ] . (b) Manual [ 133 ] and inverse-designed photonic tensor core circuit [ 129 ] .
Fig. 8 : Challenges of photonic inverse design.
Fig. 9 : Representative PIC layout automation research milestones and trends.
More in this topic

Photonic Computing

Photonic computing uses light and optical devices to perform computation, aiming for high throughput and lower energy in machine learning and signal processing.

See all 47 papers

Explore similar work

CardsList
May 20, 2026

Artificial Intelligence Reshapes Microwave Photonics

As a rapidly emerging interdisciplinary field that intrinsically integrates microwave and photonics, microwave photonics (MWP) provides disruptive solutions to overcome the fundamental bandwidth of conventional electronic systems. By exploiting the inherently ultra-wide bandwidth and low-loss characteristics of photonic technologies, MWP enables the generation, transmission, processing, and detection of microwave, millimeter-wave, and terahertz signals. Representative breakthroughs include fully photonic microwave radar systems, photonic analog-to-digital converters with bandwidth up to 320 GHz, and photonic wireless communication systems achieving data rate as high as 616 Gbit/s. Meanwhile, the rapid growth of artificial intelligence (AI) is reshaping scientific research, engineering, and daily life in unprecedented ways, such as AI for science/engineering and AI co-scientist/assistant. Correspondingly, AI is profoundly reshaping MWP in all aspects, ranging from signal generation, transmission to signal processing and detection. AI has revolutionized the design, simulation, fabrication, testing, deployment, and maintenance of MWP systems, delivering autonomous operation and exceptional efficiency beyond traditional systems. Motivated by these developments, this Review Paper provides the first comprehensive overview of AI-enabled MWP, systematically summarizing the state-of-the-art advances and presenting insights for both the academic community and the broader public.

Peng Li, Xihua Zou, Jia Ye +2

Center for Information Photonics and Communications, School of Information Science and Technology, Southwest Jiaotong University, Chengdu 611756, China · Key Laboratory of Photonic-Electronic Integration and Communication-Sensing Convergence, Southwest Jiaotong University, Chengdu 611756, China

AI for Science
AskPDF
Apr 16, 2026

Towards Topology-Aware Very Large-Scale Photonic AI Accelerators

The rapid growth of deep neural networks (DNNs) has exposed fundamental limitations in electronic accelerators, where data movement dominates energy consumption, commonly referred to as the memory wall. Photonic accelerators offer a compelling alternative due to their inherent parallelism and high-speed matrix operations. However, existing research largely focuses on device-level innovations, leaving system-level scalability insufficiently explored. In this paper, we present a scalable photonic accelerator architecture based on a modular scale-out paradigm using 4 X 4 photonic tensor core units. We perform a systematic architectural analysis that incorporates the practical scaling limits of photonic hardware, including insertion loss, fanout penalties, and laser power limits, which restrict monolithic photonic scaling. Through evaluation on representative DNN workloads (GoogleNet, ResNet-18, MobileNet, and AlphaGo Zero) with up to 1024 processing elements, we identify a topology-dominated scaling bottleneck in the photonic domain, termed the Utilization Wall, where performance is governed by grid topology rather than hardware size. We further establish the Symmetric Grid Rule, demonstrating that symmetric topologies improve utilization by up to 6X while reducing memory access by over 40% compared to linear configurations, which reveal that topology-aware scaling is essential for achieving energy-efficient and high-performance photonic AI accelerators.

Belal Jahannia, Abdolah Amirany, Hamed Dalir

Department of Electrical and Computer Engineering, University of Florida, Gainesville, FL, USA

Efficient Neural Network InferencePhotonic Computing
AskPDF
Mar 25, 2026

Machine vision with small numbers of detected photons per inference

Machine vision, including object recognition and image reconstruction, is a central technology in many consumer devices and scientific instruments. The design of machine-vision systems has been revolutionized by the adoption of end-to-end optimization, in which the optical front end and the post-processing back end are jointly optimized. However, while machine vision currently works extremely well in moderate-light or bright-light situations -- where a camera may detect thousands of photons per pixel and billions of photons per frame -- it is far more challenging in very low-light situations. We introduce photon-aware neuromorphic sensing (PANS), an approach for end-to-end optimization in highly photon-starved scenarios. The training incorporates knowledge of the low photon budget and the stochastic nature of light detection when the average number of photons per pixel is near or less than 1. We report a proof-of-principle experimental demonstration in which we performed low-light image classification using PANS, achieving 73% (82%) accuracy on FashionMNIST with an average of only 4.9 (17) detected photons in total per inference, and 86% (97%) on MNIST with 8.6 (29) detected photons -- orders of magnitude more photon-efficient than conventional approaches. We also report simulation studies showing how PANS could be applied to other classification, event-detection, and image-reconstruction tasks. By taking into account the statistics of measurement results for non-classical states or alternative sensing hardware, PANS could in principle be adapted to enable high-accuracy results in quantum and other photon-starved setups.

Shi-Yuan Ma, Jérémie Laydevant, Mandar M. Sohoni +3

School of Applied and Engineering Physics, Cornell University, Ithaca, NY 14853, USA · Present address: Research Laboratory of Electronics, Massachusetts Institute of Technology, Cambridge, MA 02139, USA. · NTT Physics and Informatics Laboratories, NTT Research, Inc., Sunnyvale, CA 94085, USA +2

Image ClassificationHardware-Software Co-DesignNeuromorphic Computing
AskPDF