Physics-guided and fabrication-aware inverse design of photonic devices using diffusion models

TL;DR

This work addresses the difficulty of designing fabrication-ready, free-form photonic devices by introducing AdjointDiffusion, a physics-guided diffusion framework that injects adjoint sensitivity gradients into the reverse-diffusion sampling process. The method learns a fabrication-aware prior from synthetic binary data and uses the adjoint FoM gradient, mapped through the diffusion Jacobian, to steer generation toward high-performance layouts without heavy post-processing. It achieves superior or competitive figures of merit with roughly $2\times 10^{2}$ simulations—orders of magnitude fewer than typical deep-learning approaches—while improving manufacturability on bent waveguides and CMOS color routers. The approach reduces reliance on complex binarization schedules and yields designs naturally aligned with fabrication constraints, with an open-source implementation to accelerate adoption in photonic inverse design.

Abstract

Designing free-form photonic devices is fundamentally challenging due to the vast number of possible geometries and the complex requirements of fabrication constraints. Traditional inverse-design approaches--whether driven by human intuition, global optimization, or adjoint-based gradient methods--often involve intricate binarization and filtering steps, while recent deep learning strategies demand prohibitively large numbers of simulations (10^5 to 10^6). To overcome these limitations, we present AdjointDiffusion, a physics-guided framework that integrates adjoint sensitivity gradients into the sampling process of diffusion models. AdjointDiffusion begins by training a diffusion network on a synthetic, fabrication-aware dataset of binary masks. During inference, we compute the adjoint gradient of a candidate structure and inject this physics-based guidance at each denoising step, steering the generative process toward high figure-of-merit (FoM) solutions without additional post-processing. We demonstrate our method on two canonical photonic design problems--a bent waveguide and a CMOS image sensor color router--and show that our method consistently outperforms state-of-the-art nonlinear optimizers (such as MMA and SLSQP) in both efficiency and manufacturability, while using orders of magnitude fewer simulations (approximately 2 x 10^2) than pure deep learning approaches (approximately 10^5 to 10^6). By eliminating complex binarization schedules and minimizing simulation overhead, AdjointDiffusion offers a streamlined, simulation-efficient, and fabrication-aware pipeline for next-generation photonic device design. Our open-source implementation is available at https://github.com/dongjin-seo2020/AdjointDiffusion.

Figures (7)

• Figure 1: Illustration of the process of integrating adjoint optimization with diffusion models for generating structures. (a) The forward and reverse diffusion process. The forward process (red arrows) starts from binary images ($\hat{\mathbf{\varepsilon}}_0$) and adds Gaussian noise until the images are completely noisy. The reverse process (blue arrows) denoises the noisy images step-by-step, using a learned model $\theta$ to reconstruct the binary images. (b) Schematics of adjoint sensitivity analysis with a two-dimensional lens. The adjoint gradient $\boldsymbol{g}_t$ is calculated for the denoised prediction $\hat{\boldsymbol{\varepsilon}}_0(\boldsymbol{\varepsilon}_t)$ by the component-wise product of direct and adjoint fields. (c) The calculated gradient from (b) is added component-wise to the generated structures in the reverse process, where conditional parameters $\sigma$ and $t$ are applied.
• Figure 2: Schematic of the problem setup for photonic inverse design. (a) Waveguide configuration with a design region of 64 $\times$ 64 pixels degree of freedom. The white arrow represents the mono-wavelength source, while the dotted lines indicate monitors. (b) CMOS image sensor (CIS) configuration with a design region of 64 $\times$ 64 pixels degree of freedom. Bloch boundary conditions on the left and right, and a PML at the top. Colored circles indicate monitor locations for each color (R,G,B).
• Figure 3: Optimization trajectories and generated structures of waveguide designs under the same fabrication conditions. (a) Bending efficiency over 100 steps of the optimization process for each algorithm. For AdjointDiffusion, the red line denotes the mean efficiency across three experimental seeds, with the shaded region indicating standard deviation. The efficiency after post-processing is marked with a star symbol. (b) Generated structures with corresponding characteristics, where Minimum Feature Size (MFS) components are highlighted in red.
• Figure 4: Maximum FoM (conversion efficiency) of generated waveguide structures across varying feature lengths (FL) and optimization steps. Three distinct feature length values are investigated: 0.224 (top row), 0.562 (middle row), and 0.895 (bottom row). Note that GA refers to the standard Gradient Ascent algorithm.
• Figure 5: Optimization trajectories and generated structures of color router. The settings and baseline approaches are identical to those in Figure \ref{['fig:wg']}, except that the optimization is performed for CMOS image sensors instead of waveguide structures.