Ultra-fast and accurate force spectrum prediction and inverse design of light-driven microstructure by deep learning

Abstract

Light can mechanically manipulate micro-/nano-particles. Recently, there has been an increasing interest in designing particles that experience controlled optical forces by tailoring light scattering. However, the huge parameter space makes traditional computational approaches impractical. Here, using data calculated from the state-of-the-art Mie scattering-Maxwell stress tensor method, deep neural networks (DNNs) are trained to study the optical forces acting on microstructures composed of a 5 × 5 square grid where each site is either empty or occupied by a dielectric sphere. Different structure configurations can tailor light scattering and forces. This paper aims to obtain a configuration that experiences different predefined forces when illuminated by light of different frequencies. The design targets are imprinted in a pseudo-optical force spectrum using a generative network. Then, by integrating all the proposed DNNs, inverse design is performed, where from a given pseudo-optical force spectrum, a microstructure satisfying the design targets is obtained. Compared to traditional approaches, the DNNs approach is several orders of magnitude faster while maintaining a high accuracy. Furthermore, for designing microstructures, this circumvents the need for iterative optimization. This approach paves the way for efficiently developing light-driven machines such as nano-drones or nano-vehicles, where tailored multiple-frequency responses are required.