Machine Learning Methods for Freeform Metamaterial Design

We will discuss computational algorithms based on deep neural networks that can accelerate the design and simulation of nanophotonic devices, using metasurfaces and metamaterials as a model system. We will discuss the use of generative networks to perform population-based optimization and elucidate how the neural network architecture can be tailored to effectively perform freeform optimization. We will also discuss how physics-augmented deep networks can be trained with a combination of data and physics constraints to serve as accurate surrogate electromagnetic solvers. A principal challenge involves configuring the algorithms in a manner that enables application to a wide range of problems, and we show how these concepts can generalize to the simulation and optimization of photonic devices involving a range of domain sizes, fitting parameters, and functions. Together, these algorithms can effectively search for the global optimum three to four orders of magnitude faster than with conventional methods. We anticipate that with proper co-design of the neural network architecture with the scientific computing task, our surrogate solver and optimizer concepts can be adapted to large scale three-dimensional photonic systems.