Optical trapping force exploration on Mie particles for structured light via multivariate sensitivity analysis and deep learning

Abstract

Optical tweezers play a critical role in exploring light–matter interactions. However, characterization of optical forces – especially for structured light fields such as circular Airy beams – remains computationally intensive under conventional numerical frameworks like the Generalized Lorenz-Mie Theory (GLMT). To overcome this limitation, we present an enhanced hybrid deep learning framework that integrates deep neural networks, GLMT simulations, and multivariate sensitivity analysis for rapid and precise prediction of optical trapping forces acting on Mie particles. Our model is trained on 5000 GLMT-calculated datasets and achieves 97.9% prediction accuracy while reducing inference latency to just 32 ms—representing a $3.9\times 10^4$-fold acceleration over traditional methods. By incorporating Sobol global sensitivity analysis, we identify the dominant beam parameters influencing optical force responses and guide the design of a fully connected neural network that achieves a mean squared error (MSE) of $2.04\times 10^{-2}$ and a relative prediction error below 1.2% for all validation data. These results demonstrate that spatial distribution variations contribute most significantly to optical force modulation and deep learning accelerates optical capability exploration, providing a deeper understanding on the optical trapping force of structured light and paving a way to quick intelligent design and optimization of structured-light-based optical tweezers.