A Neuro-symbolic Approach to Inverse Design of Thin-layer Metamaterials Under Layout Constraints

Carlo Adornetto, Marco Duca, Francesco Gallo, Gianluigi Greco, Antonio Ielo, Francesco Ricca, Simone Ventrici

Keywords: Neurosymbolic AI, Deep Learning, Inverse Design

Abstract: Inverse design aims to compute physical structures that exhibit desired properties. A prominent application in Photonics is the inverse design of metamaterials, which are artificial composite structures created by stacking layers of different materials to achieve targeted optical responses. However, designers are often interested not only in achieving target properties but also in ensuring that the resulting metamaterials satisfy specific layout constraints. Although many Deep Learning (DL) approaches have recently been proposed for inverse design, they generally fail to incorporate such constraints into the design process. In this paper, we propose a neuro-symbolic approach that combines DL-based inverse design methods with Semantic Loss to inject layout requirements into the inverse design process. Our experiments demonstrate that the proposed approach enables state-of-the-art inverse design techniques to comply with a variety of constraints inspired by the Photonics literature.

Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)

Submission Number: 12553