Enhanced Photonic Chip Design via Interpretable Machine Learning Techniques

TL;DR

This work tackles the opacity of inverse-designed photonic chips by applying Local Interpretable Model-Agnostic Explanations (LIME) to mode (de)multiplexers, revealing which geometric features influence bandwidth. By linking LIME heatmaps to targeted design changes and selecting LIME-informed initial conditions, the authors achieve substantially improved $0.5$ dB bandwidths (>$200$ nm) while maintaining reasonable transmission. The study demonstrates that interpretability tools can actively enhance inverse-design workflows, guiding initial conditions and parameter selections to produce better-performing photonic components. The approach is validated via FDTD simulations that corroborate the interpretability findings and highlights a path toward more transparent, efficient, and scalable photonic chip design with potential extension to other nanophotonic components. The practical impact lies in delivering more reliable design guidance, reducing computational cycles, and facilitating fabrication-ready inverse-designed photonic devices.

Abstract

Photonic chip design has seen significant advancements with the adoption of inverse design methodologies, offering flexibility and efficiency in optimizing device performance. However, the black-box nature of the optimization approaches, such as those used in inverse design in order to minimize a loss function or maximize coupling efficiency, poses challenges in understanding the outputs. This challenge is prevalent in machine learning-based optimization methods, which can suffer from the same lack of transparency. To this end, interpretability techniques address the opacity of optimization models. In this work, we apply interpretability techniques from machine learning, with the aim of gaining understanding of inverse design optimization used in designing photonic components, specifically two-mode multiplexers. We base our methodology on the widespread interpretability technique known as local interpretable model-agnostic explanations, or LIME. As a result, LIME-informed insights point us to more effective initial conditions, directly improving device performance. This demonstrates that interpretability methods can do more than explain models -- they can actively guide and enhance the inverse-designed photonic components. Our results demonstrate the ability of interpretable techniques to reveal underlying patterns in the inverse design process, leading to the development of better-performing components.

Figures (10)

• Figure 1: Illustration of the LIME process integrated in the interpretation of inverse design outputs. Dataset is a collection of inverse design generated samples. Neural network represents a non-linear often complex model architecture which outputs a certain metric of choice, such as prediction accuracy. Function surrogate model is a simpler model, such as a type of regression applied locally on the level of a data sample. The output of the surrogate model provides insights into the predictions of the neural network.
• Figure 2: a) Schematic representation of a photonic inverse design simulation. Design region is characterized by a permittivity distribution $\epsilon(p)$. The system features two input channels modeled as eigenmode sources (wg$_{0}$(TE$0$) and wg$_{1}$(TE$0$) modes), and one output channel (supporting wg$_{2}$(TE$0$) and wg$_{2}$(TE$1$) modes). The simulation boundaries use perfectly matched layer (PML) conditions to minimize reflection. An output field monitor evaluates performance at each iteration. (b) A mode multiplexer at various stages of the inverse design optimization. The objective function prioritized maximizing transmissions from wg$_{0}$(TE$0$) to wg$_{2}$(TE$0$) and from wg$_{1}$(TE$0$) to wg$_{2}$(TE$1$) at the wavelength of $1.55~\mu$m. The transmissions improving over the course of optimization is shown in plot (c).
• Figure 3: Inverse designed multiplexers M1-M4 in grayscale (a-d), and their transmission spectra for Mode 1 (e) and Mode 2 (f). The dotted lines show the corresponding $0.5$ dB bandwidth cut-offs. M3 and M4 have much higher bandwidths even though similar design parameters were used.
• Figure 4: Architecture of the CNN employed in this study. The network accepts grayscale images of size $256 \times 256$ as input. The model comprises three convolutional layers with filter sizes of $3 \times 3$ and increasing depths of $32$, $64$, and $128$, each followed by batch normalization (BN) and $2 \times 2$ max pooling. A spatial dropout (SpaDrop) layer with a dropout rate of $0.3$ is applied after the first max pooling layer to promote regularization. The convolutional stack is followed by a global average pooling (GAP) layer. The resulting feature vector is passed through two fully connected (dense) layers of $128$ and $32$ units, respectively, with dropout rates of $0.4$ and $0.3$. Finally, a single output node with a sigmoid activation function produces the prediction.
• Figure 5: Training (a) and Loss (b) curves of the CNN model across the number of epochs.