Hybrid neural network-powered inverse design for implementing a highly efficient multimode interference coupler

Multimode interference couplers have been increasingly utilized in a variety of photonic scenarios. However, manually designing a coupler with sufficiently high coupling efficiency demands substantial time and effort, without guaranteeing the required performance. In this study, we present a machine learning-driven methodology that relies on a hybrid neural network and the Nelder-Mead algorithm to inversely design an efficient 1 × 4 coupler. The maximum output power values and corresponding coupler parameters are automatically and swiftly deduced through multiple iterations. The optimized average coupling efficiency, insertion loss, and power imbalance are −6.05 dB, 0.033 dB, and 0.039 dB, respectively, over the telecommunication spectral band spanning 1530–1630 nm. In comparison to conventional design methods, our approach significantly diminishes the insertion loss by approximately 0.04–0.1 dB. Our scheme shows promise in propelling and simplifying the designs of diverse types of couplers.