Abstract:
We show that design optimizations, an integral but time-consuming component of optical engineering, can be significantly sped-up when paired with deep neural networks (DNNs). By using the DNN indirectly for choosing initializations and candidate preselection, our approach obviates the need for large networks, big datasets, long training epochs, and excessive hyperparameter optimization. For a 16-layered thin-film design problem, our surrogate-assisted differential evolution (DE) algorithm is able to achieve similar optimal solutions as that of an unassisted DE using only 10% of the function evaluation budget. Our approach is a promising option for the optimal design of optical devices and systems.