Massively parallel pixel-by-pixel nanophotonic optimization using a Green's function formalism
We introduce an efficient parallelization scheme to implement pixel-by-pixel nanophotonic optimization using a Green's function based formalism. The crucial insight in our proposal is the reframing of the optimization algorithm as a large-scale data processing pipeline, which allows for the efficient distribution of computational tasks across thousands of workers. We demonstrate the utility of our implementation by exercising it to optimize a high numerical aperture focusing metalens at problem sizes that would otherwise be far out of reach for the Green's function based method. Finally, we highlight the connection to powerful ideas from reinforcement learning as a natural corollary of reinterpreting the nanophotonic inverse design problem as a graph traversal enabled by the pixel-by-pixel optimization paradigm.