Deep learning-driven inverse design of all-dielectric silicon metasurfaces with targeted spectral response

Abstract

The inverse design of all-dielectric metasurfaces faces challenges, including the reliance on full-spectrum input and the limitation of structure design to predefined geometries. This paper constructs a deep-learning-driven inverse-design framework that combines a fitness function with multi-strategy optimization algorithms to achieve the on-demand design of all-silicon metasurfaces. Here, the fitness function enables the quantification of convex-shaped spectral peaks (based on their wavelength and intensity) by combining a weight mechanism and avoiding interference from irrelevant wavelength bands. Meanwhile, the target error is monitored in real time through the fitness function value, enabling precise control over single-peak, dual-peak, and three-peak spectral responses. Moreover, three complementary optimization algorithms are considered: discrete optimization to explore full-degree-of-freedom structures, linear optimization to generate block-shaped and processable structures through parameterized row scanning, and shape optimization to simplify the process by restricting structures to regular geometries. The effective optimization of single-peak, dual-peak, and three-peak from the numerical simulation confirms that our proposed framework possesses distinct advantages, including a minimal target wavelength error, an order-of-magnitude higher efficiency, and excellent consistency between the fitness function value and the actual spectral deviation. This work enables, for the first time, the transformation of the silicon metasurface design from a “full-spectrum trial-and-error” approach to a “key-feature-driven reverse engineering” paradigm.