Physics-Informed Neural Network for Cross-Domain Predictive Control of Tapered Amplifier Thermal Stabilization

Thermally induced laser noise critically limits the sensitivity of quantum sensor arrays using ultra-stable amplified lasers, stemming from nonlinear gain-temperature coupling in tapered amplifiers (TAs). Existing temperature control approaches exhibit limited performance across varying laser power regimes, resulting in either measurement sensitivity degradation or demanding recalibration procedures in quantum precision measurements. This work presents an innovative architecture integrating a specialized physics-informed gated recurrent unit (PI-GRU) network with a model predictive control (MPC) strategy, achieving high stability of TA temperature across multi-power laser operations. By embedding physical soft constraints, we establish a predictive model with enhanced physical consistency that enables reliable extrapolation capability beyond trained operational scenarios. Based on the model’s accurate multi-step predictions, a hierarchical parallel MPC architecture is constructed to achieve real-time compensation of thermal instability. While trained exclusively on safe low-power samples, the proposed PI-GRU MPC demonstrates significant improvements at high-power operations, achieving an over 58.2% greater prediction accuracy and an over 50.0% better temperature stability compared to data-driven methods, as experimentally validated. The synchronization of physics-informed neural networks with MPC frameworks offers a novel solution to robustness challenges in cross-domain control under varying operating regimes.