Towards Explainable Inverse Design for Photonics via Integrated Gradients

TL;DR

The paper tackles the interpretability gap in adjoint-based inverse photonic design by introducing an end-to-end workflow that pairs SPINS-B generated WDM layouts with a lightweight CNN surrogate and pixel-level Integrated Gradients to reveal which regions drive performance metrics. IG saliency consistently localizes to physically meaningful features such as tapers and splitter hubs, providing design intuition beyond the conventional adjoint solution while preserving the original physics solver and objective. The authors release a $N=500$ design corpus, the CNN surrogate, and the IG analyses to enable reuse across photonic components. This approach enables interpretable, data-driven guidance for inverse-design workflows and can be integrated into design loops to improve yield and insight without altering the underlying physics.

Abstract

Adjoint-based inverse design yields compact, high-performance nanophotonic devices, but the mapping from pixel-level layouts to optical figures of merit remains hard to interpret. We present a simple pipeline that (i) generates a large set of wavelength demultiplexers (WDMs) with SPINS-B, (ii) records each final 2D layout and its spectral metrics (e.g., transmitted power at 1310 nm and 1550 nm), and (iii) trains a lightweight convolutional surrogate to predict these metrics from layouts, enabling (iv) gradient-based attribution via Integrated Gradients (IG) to highlight specific regions most responsible for performance. On a corpus of sampled WDMs, IG saliency consistently localizes to physically meaningful features (e.g., tapers and splitter hubs), offering design intuition that complements adjoint optimization. Our contribution is an end-to-end, data-driven workflow--SPINS-B dataset, CNN surrogate, and IG analysis--that turns inverse-designed layouts into interpretable attributions without modifying the physics solver or objective, and that can be reused for other photonic components.

Figures (6)

• Figure 1: Inverse-designed WDM: 1310nm light is routed to the right port and 1550nm light is routed to the left port.
• Figure 2: End-to-end pipeline: SPINS-B $\rightarrow$ dataset $\rightarrow$ CNN surrogate $\rightarrow$ IG $\rightarrow$ Feature Importance.
• Figure 3: Compact CNN surrogate used to regress power from a binary layout and enable pixel-wise IG.
• Figure 4: Training and validation MSE loss curves (values $\times10^{-4}$).
• Figure 5: Tracing of $\alpha$ along the Integrated Gradients path. IG activations accumulate progressively, highlighting regions along the $\alpha$-path.