Data-driven eigenmode Estimation of optical fibers in TMI-Regime by exploitation of physically constrained glass box machine learning model

This work presents a concise overview of the development and application of a fully interpretable machine learning (ML) model designed to identify the eigenstates of optical fibers operating in the regime of transverse mode instability (TMI). The core principle of this approach lies in exploiting the structural equivalence between the ML model and the underlying physical system, achieved through the integration of a physically motivated architecture and regularization techniques. We outline the design strategy for the trainable model, emphasizing its simplification and the associated training procedures. The light propagation formalism employed in this study leverages complex-valued representations, which are explicitly embedded into the model’s architecture and systematically addressed during the training process. The proposed model is trained using high-speed measurements of TMI behavior in a weakly guiding large-mode-area (LMA) optical fiber at an output power of 365 W. These measurements enable the discrimination of optical signals across four distinct polarization states - 0°, 90°, -45°, and σ - in accordance with the established Jones formalism. Furthermore, we detail a regularization scheme for these signals and demonstrate its alignment with the corresponding training framework.