UCLA

Designing photonic structures and obtaining optimal responses is still a difficult task. A large number of gradient based algorithms and adjoint solvers are used, and the results have been effective in that complex free form geometries of devices are capable of being created. The fact remains, however, that many solvers still tend to be black boxes, and there is very little transparency in how the solver comes to its conclusion. Furthermore, because many of these solvers are gradient based, additional knowledge of the device is needed, otherwise the solver can get trapped in a local minimum and not reach the true optimal geometry. To this end, an inverse design framework that combines adjoint optimization, automated machine learning (AutoML), and explainable artificial intelligence (XAI) is presented in order to determine both the relation between device structure and performance and also minimize the effect of local minima trapping. This framework is used in the inverse design of waveguides and achieves an average of 43\% increase in performance. Consequently, this study portrays how to extend beyond traditional inverse design solvers in terms of both performance and ability to elucidate the solvers' decisions.