Numerical Simulation Engines and Deep Learning Architectures for Inverse Design of Multilayer Metamaterials
Abstract
The inverse design of multilayer metamaterials —artificial structures with engineered electromagnetic, acoustic, thermal, or mechanical properties— represents a high-dimensional optimisation challenge that has driven the integration of deep learning into computational photonics and materials science. This paper presents a systematic literature review conducted under the PRISMA 2020 protocol, aimed at mapping the numerical simulation methods employed as forward engines or training data generators within deep learning-assisted inverse design pipelines. Starting from 392 records retrieved from Scopus and Web of Science over the 2010–2025 period, a multi-stage screening process based on a heuristic prioritisation index termed CARS (Citation-Adjusted Relevance Score) identified 44 core articles. Eleven distinct numerical methods were identified and organised into two conceptual layers: classical simulation engines (finite element method, topology optimisation, guided-mode expansion, adjoint method, and genetic algorithms) and deep learning architectures (tandem CNN, transformer, GAN, PINN, and diffractive neural networks). Results show a strong dominance of the electromagnetic-photonic domain (72%), accelerated growth of deep learning-based publications from 2022 onward, and a clear trend toward hybrid physics, data and learning pipelines. The review concludes by proposing a prioritised experimental roadmap for future research.