Digital design of mode decomposition systems for multimode fibers using physics-informed neural network

Multimode fiber is playing an increasingly important role in both classical and quantum communication. However, distortion occurring inside the fiber poses a significant challenge for practical deployment. A closed-loop control system is typically required to compensate for this distortion, which in turn necessitates accurate measurement of the light field within the fiber. Mode decomposition enables access to the exact modal amplitudes and phase weights, forming a foundation for further technological development. We propose a pretraining-based, physics-informed neural network for fast mode decomposition in multimode fiber. Using only intensity images, our approach achieves over 98% correlation in decomposing more than 1000 spatial modes. This work offers a promising path toward integrating digital optical technologies into fiber-based applications.