Deep Learning and Evolutionary Optimization Methods for Electromagnetic Devices

Metalens design and fabrication in the optical regime presents many exciting opportunities for developing state-of-the-art optics technologies. However, fabricating electrically large devices at scale for optical wavelengths is a major challenge due to limits in existing lithographic fabrication techniques. In this summary, we present our work for counteracting performance losses arising from geometric uncertainty of fabricated devices (e.g., metasurface supercells). We employed Deep Learning (DL) methods to overcome the inherent intractability of exhaustive robustness testing, and a multi-objective evolutionary algorithm was used to find tolerant yet efficient device designs. When amortized over several optimization runs, the proposed DL-augmented optimization scheme offers a speedup of more than 10 times over a full-wave-only strategy.