Machine learning design of subwavelengh integrated photonic devices

Abstract

Use of subwavelength metastructures opens new degrees of freedom to control and manipulate propagation of light in planar waveguide devices. This advantage comes with the cost of increased design complexity since more parameters must be simultaneously optimized. Here we show how machine learning dimensionality reduction can be used to obtain a compact representation of a multi-parameter design space revealing the relationship between different design parameters. This provides the designer with a global perspective on the design space and enables informed decisions based on the relative priorities of different performance metrics.