Geometric Parameter Prediction with Color Reproduction of Silicon in Reverse Design and Measurement

Abstract

The design of nanostructure colors is influenced by mechanisms such as quantum size effects, surface plasmon resonance, and structural coloration. These optical properties arise from the interaction between localized magnetic and electric dipole resonances, rendering them highly sensitive to changes in geometric parameters. However, conventional analytical methods are inefficient in optimizing geometric parameters to achieve target colors, particularly when faced with the challenges of large-scale and diverse structural color designs. To address this limitation, we propose a design framework based on a bidirectional deep neural network (DNN) consisting of both a forward network and an inverse design network. The forward network learns the relationship between geometry and color response through parameter scans, enabling precise color prediction for specific geometries. The inverse design network derives the corresponding geometry from target color coordinates (CIE1931 color space) and tackles the multi-solution challenges in inverse design by cross-validating with the forward network. Rigorous computational modeling demonstrates that this approach can generate over one million visible-spectrum nanostructure colors with a theoretically predicted color reproduction rate exceeding 98%. This research presents a highly efficient and accurate framework for the design of high-precision optical components, including those used in silicon-based color processing, optical displays, sensors, and photovoltaic systems.