Physics-informed Data-driven Model for Predicting Thermal Crosstalk in a Photonic Integrated Circuit

Programmable photonic circuits are promising platform for matrix–vector multiplication in optical computing applications. However, effects like thermal crosstalk degrade their performance and limit scalability. To address this, different data-driven and analytical modeling approaches that include crosstalk were introduced. While data-driven models capture the effect more accurately, they usually require large training sets and cannot guarantee physical plausibility. We propose a physics-informed machine learning approach for predicting thermal crosstalk in a photonic chip by introducing physics-based cost function term, hence leveraging inherent physics while benefiting from measurement data. We demonstrate how varying relative contribution of the physics-based term influences a data-driven model's behavior and training. Moreover, the physics-based term is parameterized with learnable parameters that are trained jointly with the model and remain consistent across the range of contribution levels.