Machine learning design of spectral-selective infrared metasurfaces based on Conway patterns

Abstract

Metasurfaces offer precise control over infrared radiation, holding great potential for thermal management, stealth technology, and sensing applications. However, existing computational approaches, especially machine learning-based methods, typically rely on simplistic geometries and limited material selections, constraining design complexity and performance. To overcome these challenges, we introduce a machine learning-driven platform employing Conway-inspired patterns to achieve band-selective infrared metasurfaces. Our system integrates an automated optimization framework combining intricate, rule-based Conway morphologies, standard geometric parameters, and an extensive database of 71 materials. Leveraging a deep neural network for forward predictions and particle swarm optimization for efficient inverse design, our approach successfully produces diverse, high-performance metasurfaces. These designs exhibit superior radiative cooling and stealth properties within the critical 5–8 μm atmospheric transparency window. Generated mosaic-like patterns are further optimized through image filtering and symmetry processing, enhancing fabrication feasibility and minimizing polarization dependence. This comprehensive design paradigm significantly broadens the available design space, facilitating the discovery of novel metasurface structures with multifunctional capabilities in optics and thermal management.