Physics-Based Flow Matching for Full-Field Prediction of Silicon Photonic Devices Preprint: APS/123-QED

Joseph Quaratiello 1 and Anthony Rizzo 1 · 2026

Read Physics-Based Flow Matching for Full-Field Prediction of Silicon Photonic Devices Preprint: APS/123-QED by Joseph Quaratiello 1 and Anthony Rizzo 1 (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2605-06929.txt.

Introduction

Joseph Quaratiello and Anthony Rizzo build a neural surrogate that predicts electromagnetic fields inside silicon photonic devices, replacing slow simulations during design. Simulating fields well with finite-difference time-domain methods is costly even for simple shapes. Their PIC-Flow model learns, through conditional flow matching, a speed field that carries random noise to physically valid field solutions for a given geometry and wavelength.

Training adds a Helmholtz equation residual penalty, with boundary pixels masked out where numerical errors dominate, so the physics constraint stays meaningful. The training set holds twenty-two thousand simulations of three device families. The model also generalizes to held-out device types like bends and tapers, a foundation toward fast, broad design exploration.

Catalog record

CREATORJoseph Quaratiello 1 and Anthony Rizzo 1
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE45 KB
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2605.06929

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Physics-Based Flow Matching for Full-Field Prediction of Silicon Photonic Devices Preprint: APS/123-QED Joseph Quaratiello 1 and Anthony Rizzo 1 Abstract Designing photonic integrated circuits requires accurate electromagnetic field simulations, which remain computationally expensive even for simple device geometries. We present PIC-Flow, a generative neural surrogate that predicts electromagnetic field distributions for photonic devices given their geometry and operating wavelength as an alternative to costly finite-difference time-domain (FDTD) simulations. Our approach combines three key ideas: (i) conditional flow matching as the generative framework, learning a velocity field that transports Gaussian noise to physically valid field solutions; (ii) a real-valued U-Net operating on split real and imaginary field channels; and (iii) physics-constrained training through a Helmholtz residual loss enforcing ∇ 2 E z + k 0 2 ​ ε ​ E z = 0 \nabla^{2}E_{z}+k_{0}^{2}\varepsilon E_{z}=0 . We introduce an interface-aware masking scheme for the Helmholtz residual that excludes dielectric boundary pixels where finite-difference stencil errors dominate, yielding a physically meaningful compliance metric. The data set consists of 22,500 ground-truth FDTD simulations split evenly between multimode interferometers, Y-branches, and directional couplers at λ = 1.55 ​ μ \lambda=1.55\,\mu m in an 80/10/10 split between training, validation, and test sets. We evaluate ablations on the network against the held out test devices and also show that the model generalizes to held out device classes such as S-bends, tapers, and cascaded Y-branches. Rather than a drop-in replacement for FDTD, this work establishes a foundation that, with broader data coverage, more compute, and further training optimization, could scale toward broadband, device-agnostic field prediction with dramatically improved runtime for rapid design-space exploration of complex photonic devices and circuits. I Introduction Silicon photonic devices used in photonic integrated circuits (PICs) enable compact optical systems for communications 27 ; 23 ; 6 , sensing 24 ; 32 ; 34 , and computing 1 ; 7 ; 3 , but practical design still depends on full-wave electromagnetic simulation and iterative optimization. Finite-difference time-domain (FDTD) solvers remain the main tool for accurate prediction across device geometries and wavelengths 28 ; 31 , yet repeated FDTD simulations become a bottleneck for parameter sweeps and gradient-based inverse-design loops. This has motivated machine-learning surrogates for forward prediction and learning-assisted inverse design. A key remaining challenge is accurate full-field prediction of complex electromagnetic responses: phase must be handled consistently, solutions should satisfy the wave equation, and models must generalize across device families rather than overfitting to one geometry class. Predicted fields should also preserve device-level behavior such as port-to-port transmission, reflection, and splitting ratios, the quantities ultimately used in photonic circuit design. We address this gap with a physics-based conditional flow-matching model for fast full-field prediction of silicon photonic devices. Building on flow matching for generative modeling 12 and physics-based flow matching for PDE-constrained generation 2 , we train a real-valued U-Net to generate the real and imaginary components of the electromagnetic field. Physics is enforced through a Helmholtz residual loss with an interface-aware mask that excludes dielectric boundary pixels where finite-difference stencil errors dominate, yielding a meaningful compliance metric. We study three matched training objectives: flow matching only, flow matching plus phase loss, and flow matching plus both phase and Helmholtz residual losses. This isolates how data fidelity, phase alignment, and PDE compliance affect full-field accuracy. We evaluate the approach on multimode interferometers, Y-branches, and directional couplers, producing full-field predictions in a single forward pass significantly faster than standard FDTD simulations. This work makes four main contributions: (i) a flow-matching neural surrogate that predicts the complex E_{z} field of parameterized silicon photonic devices conditioned on permittivity, source mask, and wavelength; (ii) a masked Helmholtz residual loss that excludes PML, source, and dielectric-interface pixels, defining a physically meaningful per-sample compliance metric ( \rho_{R} , Eq. ( 8 )); (iii) an ablation showing that combining flow matching with phase and Helmholtz residual losses improves wave-equation compliance on held-out test samples without sacrificing reconstruction quality, together with a wall-clock benchmark against multi-threaded CPU FDTD on the same compute node; and (iv) evidence that the trained model generalizes to held-out device classes (S-bends, long tapers, cascaded 1\times 3 Y-branches) that were intentionally omitted from the training data set. This work establishes a proof-of-principle foundation that can dramatically accelerate design cycles for PIC development relative to current standards. II Related Work Our work connects three lines of research in the domain of computational electromagnetics and machine learning: learning-based surrogates for electromagnetic simulation, physics-informed or physics-constrained training for PDE-governed systems, and generative models for scientific computing. In this section, we provide a brief survey of the current state-of-the-art and discuss the novel contributions of PIC-Flow relative to previous demonstrations in each domain. We additionally briefly discuss inverse design and its relationship to PIC-Flow. Neural surrogates for photonic simulation. Deep learning has been used to approximate forward electromagnetic responses and accelerate design-space exploration across many photonic structures. Jiang et al. survey deep neural networks for evaluation and design, covering supervised surrogates, inverse models, and hybrid workflows 9 ; Ma et al. provide a broader review of deep learning for photonic structure design 15 . In integrated photonics, Tahersima et al. demonstrate data-driven inverse design of compact power splitters 29 , and Tang et al. introduce a generative model for inverse design of nanophotonic components 30 . Chen et al. train a physics-augmented U-Net 25 (WaveY-Net) to predict near-fields of freeform nanophotonic devices, using Maxwell’s equations as part of the training loss 5 , and Lim and Psaltis use Maxwell’s equations directly as the sole training objective for a physics-driven DNN 11 . These works show the value of learned surrogates, but they also highlight the need for models that predict full-field complex solutions, not only scalar figures of merit, and that generalize across device families and wavelengths. Physics-informed and physics-constrained training. Physics-informed neural networks (PINNs) offer a general way to enforce PDE residuals in the loss for forward and inverse problems 21 ; in photonics this often means wave-equation residuals, such as Helmholtz, or boundary and energy constraints. However, PINNs typically optimize per-instance and do not amortize across a family of geometries. Neural operators such as the Fourier Neural Operator 10 and DeepONet 14 learn mappings between function spaces and have been applied to PDE problems at scale, but the photonic setting still requires accurate amplitude and phase prediction because interference effects determine functionality. Our approach uses a Helmholtz residual penalty on predicted complex fields as a lightweight, architecture-agnostic physics constraint that amortizes across thousands of geometries. Generative models for PDE-governed systems. Generative models can represent distributions over fields and enable conditional sampling for design. Score-based diffusion models 26 have been applied to PDE-governed problems: Huang et al. use diffusion-based generation to solve Helmholtz and other equations under partial observation 8 . Flow matching 12 provides a simulation-free alternative by learning a time-dependent velocity field along a probability path, offering straighter transport paths and simpler training than diffusion. Baldan et al. extend this idea to physics-constrained generation with physics-based flow matching (PBFM), integrating PDE residuals during training 2 . Our work is closest to PBFM, but targets electromagnetic fields and silicon photonic device families, combining a U-Net field generator with a Helmholtz residual constraint and phase-aware supervision. Inverse design. Inverse

Read the complete text