| Abstract | Integrated photonic devices are at the basis of modern optical communications but their design remains a time-consuming and mostly iterative process. Recent design methods employing optimization algorithms make it possible to simultaneously optimize a large number of parameters and generate non-trivial geometries and novel functionalities. However, these approaches give little insight on the influence of the design variables on device performance, nor address the possibility of degenerate designs with similar performance. Using machine learning methods, we demonstrate through the design of a vertical grating coupler that a large number of degenerate designs with regard to the primary objective (coupling efficiency) exist, but they have distinct properties in other performance criteria. Pattern recognition reveals the relationship between these designs and reduces the characterising variables from the original five to two. This finding enables the exhaustive yet efficient coverage of the design space, being mapped onto a 2-dimensional hyperplane. The interplay of the design parameters and multiple performance criteria then can be clearly visualized with a vastly reduced computational effort, allowing the designer to understand and balance competing design requirements. We believe this is the first time such a global perspective is presented for high dimensional design problems in integrated photonics, representing a major shift in how modern photonic design is approached. |
|---|
