Deep-learning-based model for the inverse design of a reconfigurable multi-tap optical filter

Integrated and reconfigurable optical filters have been widely used in modern communication and signal processing systems, and the inverse design of them remains an important problem. In this study, we introduce an anti-non-uniqueness deep learning network for the inverse design of a multi-tap reconfigurable optical filter. By employing eight optical switches configuration between cross and bar status, the integrated tunable optical delay line can act as a reconfigurable multi-tap optical filter with at most 256 taps. For its inverse design, the proposed deep learning model aims to figure out the relationship between control parameters and spectrum features of the optical filter. An improved tandem neural network has been developed with its loss function comprehensively optimized. Experiments have proved that the model helps making precise predictions regardless of the multi-solution challenge of this optical filter's inverse design, facilitating its real application in practice.