University of Ottawa

Published purpose

Simulating optical components at micro and nanoscale using general purpose 3D optical solvers is often computationally expensive. The computational burden can be reduced significantly if structures in question can be modelled well in 2D. For most nanophotonic structures, however, current approaches to 2D modelling are inaccurate at best and, at worst, completely fail to give even qualitative design guidance. Previous work has demonstrated that machine learning models can create a reliable link between 2D and 3D representations of select grating coupler configurations that include metamaterials. This project expands on that work by offering to build a single joint surrogate model representing multiple grating coupler configurations simultaneously and leverage active learning tools to perform efficient data collection to train such a complex model. The use of the resulting model will be demonstrated in the inverse design of highperformance couplers with improved fabrication robustness and novel dispersion properties