Machine learning of all-dielectric core–shell nanostructures: the critical role of the objective function in inverse design

Abstract

To integrate nanophotonics into light-based technologies, it is critical to elicit a desired optical response from its fundamental component, a nanoresonator. Because the optical resonance of a nanoresonator depends strongly on its base material and structural features, machine learning has been contemplated to enhance the design and optimization processes. However, its accuracy in searching the vast parameter space of nanophotonics still poses unresolved questions. Here, we show how the choice of objective functions, in combination with trained neural networks, can drastically change the optimization process—even for a simple nanophotonic structure. To assess how different objective functions select the correct structural parameters that generate a desired optical Mie response, we use a simple core–shell, all-dielectric nanostructure as the benchmark. By controlling the proportion of training data, which represents the “experience” level, we also quantify how the various objective functions perform in finding the ground-truth parameters. Our findings demonstrate that certain objective functions exhibit improved accuracy when used with highly “experienced” neural networks. Surprisingly, we also find other objective functions that perform better when paired with less “experienced” neural networks. Taken together, our results emphasize that it is critical to understand how neural networks are coupled to optimization schemes, as is evident even when a simple core–shell nanostructure is used.