Cycle cGAN-driven inverse design of 1 × 2 MMI couplers with arbitrary power-splitting ratios and high efficiency

Multimode interference (MMI) couplers are extensively used as optical power splitters in photonic integrated circuits due to their compact size, broadband performance, and robust fabrication tolerance. However, achieving arbitrary power-splitting ratios with low excess losses (ELs) presents considerable challenges and often necessitates substantial manual adjustments and high computational costs. This study introduces a cycle-consistent conditional generative adversarial network (cycle cGAN) to address these limitations. The proposed framework integrates a cGAN as an inverse design model with a fully connected neural network serving as a forward predictive model for cycle-consistency verification. Applied to the design of 1 × 2 MMI couplers operating at a wavelength of 1550 nm, the cycle cGAN efficiently generates designs with power-splitting ratios ranging from 1:99 to 99:1 while maintaining ELs below 0.7 dB. This method substantially reduces computational demands and ensures physically feasible designs, offering a promising solution for advancing silicon photonics and photonic integration.