Automated Design of Photonic Circuits Utilizing Deep Learning

Traditional photonic design methods utilizing neural networks achieve the mapping between design parameters of specific optical devices and output responses. However, the connection structure design of photonic circuits composed of multiple optical elements remains underexplored. Therefore, we have proposed a novel and effective parametric matrix representation that is capable of realizing the mapping from different connection structures, composed of given optical elements with fixed physical parameters, to their responses. Furthermore, our parameterization approach is compatible with various kinds of connection structures, including different numbers of component devices and distinct topologies. The comparative test of forward models, trained in the same network architecture but using different input parameter representations, has demonstrated the superior generalization of our approach. Leveraging the well-performing forward model, we have implemented inverse design of connection structures using both deep neural networks and evolutionary algorithms, which can effectively find suitable structures for target responses.