Sample-efficient deep learning for accelerating photonic inverse design

Abstract

Data-driven techniques like deep learning (DL) are currently being explored for inverse design problems in photonics (especially nanophotonics) to deal with the vast search space of materials and nanostructures. Many challenges need to be overcome to fully realize the potential of this approach; current workflows are specific to predefined shapes and require large upfront investments in dataset creation and model hyperparameter search. We report an improved workflow for DL based acceleration of evolutionary optimizations for scenarios where past simulation data is nonexistent or highly inadequate and demonstrate its utility considering the example problem of multilayered thin-film optics design. For obtaining sample-efficiency in surrogate training, novel training loss functions that emphasize a model's ability to predict a structurally similar spectral response rather than minimizing local approximation error are proposed. The workflow is of interest to extend the ambit of DL based optics design to complicated structures whose spectra are computationally expensive to calculate.