Abstract
Predicting the effects of physical perturbations on optical channels iscritical for advanced photonic devices, but existing modelling techniques areoften computationally intensive or require exhaustive characterisation. Wepresent a novel data-efficient machine learning framework that learns theperturbation-dependent transmission matrix of a multimode fibre. To overcomethe challenge of modelling the resulting highly oscillatory functions, weencode the perturbation into a Fourier Feature basis, enabling a compactmulti-layer perceptron to learn the mapping with high fidelity. On experimentaldata from a compressed fibre, our model predicts the output field with a 0.995complex correlation to the ground truth, improving accuracy by an order ofmagnitude over standard networks while using 85\% fewer parameters. Thisapproach provides a general tool for modelling complex optical systems fromsparse measurements.