Machine learning-driven prediction and inverse design of optical forces near gradient metasurfaces

Due to high level of control over local field distribution, metasurfaces can potentially provide an attractive platform for creating and controlling customized optical traps for nanoscale objects. On the other hand, the complexity of wave-matter interactions in the near-field increases the computational cost associated with analyzing these structures and, more critically, hinders efficient inverse design of desired optical forces near such structures. In this work, we introduce a machine learning-based method that circumvents both limitations. Employing tandem neural network configuration, we demonstrate a surrogate solver for calculating local optical forces at various distances across a gradient metasurface. Next, we showed that, with proper training, our model can inverse design metastructures to implement prescribed forces either across the entire surface or locally defined over a specific region. Our data-driven approach and novel platform provides a valuable tool for the design and analysis of new optical nanotweezers with implications for applications relying on near-field engineering.