PhoTOS: Topology Optimization of Photonic Components using a Shape Library

TL;DR

PhoTOS addresses the fabrication-aware topology optimization of photonic components by combining a discrete shape library with a differentiable, latent-space representation learned by a Convo-implicit VAE. Shapes are transformed and assembled in the design domain, mapped to density and permittivity fields via a geometric projection, and optimized under performance objectives (scattering spectra) and fabrication constraints ($MFS$ and $MSD$) using gradient-based MMA with AD-based sensitivities. Key contributions include extending feature-mapping to multiple shapes, introducing the Convo-implicit VAE for continuous shape representations, and enforcing latent-space and separation constraints to ensure library-consistent designs, demonstrated on waveguide bends and mode converters. The framework yields fabricable, high-performance photonic components and opens paths for 3D extensions, larger shape libraries, and integration with more advanced fabrication models.

Abstract

Topology Optimization (TO) holds the promise of designing next-generation compact and efficient photonic components. However, ensuring the optimized designs comply with fabrication constraints imposed by semiconductor foundries remains a challenge. This work presents a TO framework that guarantees designs satisfy fabrication criteria, particularly minimum feature size and separation. Leveraging recent advancements in machine learning and feature mapping methods, our approach constructs components by transforming shapes from a predefined library, simplifying constraint enforcement. Specifically, we introduce a Convo-implicit Variational Autoencoder to encode the discrete shape library into a differentiable space, enabling gradient-based optimization. The efficacy of our framework is demonstrated through the design of several common photonic components.

Figures (13)

• Figure 1: Graphical abstract: Given a predefined library of shapes, a Convo-implicit Variational Autoencoder (VAE) is trained to encode them in a differentiable latent space. Shape instances are then selected from this latent space and subjected to rotation, scaling, and translation to populate the design space, yielding an optimized photonic component.
• Figure 2: (a) Photonic device design domain and boundary conditions. (b) Density-based topology optimization. (c) Topology optimization using single shape-based feature mapping. (d) Topology optimization using multiple generic shapes.
• Figure 3: A library of $n_L \; (=15)$ prescribed shapes.
• Figure 4: Shape instances selected from the library are translated, oriented, and scaled onto the design domain.
• Figure 5: To ensure fabricability, we impose: (a) minimum feature size constraint and (b) minimum separation distance constraints.