“…For photonics, the most versatile and computationally efficient inverse design algorithms are based on freeform topology optimization, − in which every dielectric voxel in the simulation domain is a free parameter that can be modified using the adjoint variables method , or autodifferentiation. , The ability for topology optimization to explore exceptionally large design spaces without approximations to Maxwell’s equations has led to demonstrations of freeform metagratings, ,, metalenses, ,, and on-chip photonic devices with exceptional performance. However, hard minimum feature size tolerances − and robustness criteria − can be difficult to incorporate into topology optimized designs, even with heuristic methods that have been developed to address these issues, − , and there remain experimental challenges to accurately manufacturing complex freeform shapes. In addition, the operating principles of these devices lack straightforward physical interpretability.…”