Machine Learning Enables Inverse Design of Optically Driven Microscopic Metavehicles

Journal article, 2026

Lateral optical forces can propel microscopic objects across a surface even when the incident light carries no net momentum in the propagation direction. Here we explore how machine-learning-based inverse design can optimize this process for metavehicles— flat microparticles incorporating silicon metagratings that deflect light at high angles. The optimized design achieves a maximum unidirectional normal-to-lateral momentum deflection efficiency of ∼88%, representing a 6% improvement over the best fabricable training-set structure and ∼60% improvement compared to the original metavehicle structure. Optimized metavehicles exhibit lateral propulsion speeds in water significantly higher than previously reported. The results demonstrate that inverse design guided by machine learning can substantially enhance optomechanical performance, paving the way toward efficient light-driven micro-transporters and actuators for microbiology and microfluidics. In addition, our strategy yields a library of high-efficiency metagrating designs for diverse photonic components, including high-numerical-aperture axicons, metalenses, and beam deflectors.