Dimensionality reduction in designing advanced silicon photonic components

Design of modern integrated nanophotonic components requires increasingly sophisticated optical simulation and optimization tools. Modeling and computational challenges arise with the increase in the number of design parameters and the introduction of multiple and often competing performance criteria. In such high dimensional design parameter spaces, it becomes difficult to navigate, explore, and visualize the best candidate designs that satisfy all the requirements. We present our recently developed approach that leverages dimensionality reduction - an area of machine learning – to identify and efficiently investigate only the most relevant portion of the design space. Once this reduced space is found, mapping and optimization can often be achieved several orders of magnitude faster than in the original design space. We showcase our approach on several design scenarios focusing on components such as optical grating couplers and power splitters. We employ principal component analysis for linear dimensionality reduction, achieving impressive performance despite its simplicity. We also demonstrate the use of a non-linear technique, i.e. neural-network based autoencoders, which can improve the effectiveness of dimensionality reduction even further. All components have nontrivial regions of interest in their design space that are identified and explored through the evaluation of various performance metrics. Visualizations of these regions offer a global picture of device behavior. Different component geometries can then be chosen depending on specific performance requirements or fabrication constraints. The proposed framework can be easily integrated into various design toolkits.