Machine Accelerated Nano-Targeted Inhomogeneous Structures

Optical metamaterials can achieve optical performance specifications otherwise unattainable compared to their bulk constituents via the use of skillfully engineered nanoscale features. These features function as sub-wavelength optical scatterers, causing incident light to interact with metamaterials in a manner which is crucially dependent on their geometry. To design the next generation of functional materials, scientists and engineers must realize new methods which prudently explore the near-infinite number of possible metamaterial arrangements. We present a general methodology which harnesses machine learning to accelerate the process of metamaterial design and apply it specifically here, as a test case, to a silicon on insulator (SOI) reflective metasurface consisting of an array of nano-pillars. In particular, we implement artificial neural networks (ANNs) to solve the inverse design problem; i.e. we prescribe a desired reflection profile and utilize ANNs to advise what metasurface array of SOI nano-pillars gives the closest result. Utilizing a synthetic data set of devices simulated with rigorous coupled wave analysis (RCWA), we create an ANN-accelerated simulator, achieving a computational speedup of O(106) over the relatively quick RCWA simulation method. We then couple this simulator with an ANN-based predictor, which directly binds the SOI metasurfaces in question to a targeted optical performance. Together, this simulator/predictor ANNs combination provides a general framework in rapidly evaluating and designing potential optical metamaterial devices beyond what is currently possible via straightforward simulation and more standard optimization methods.