Knowledge-driven adaptive alignment method for reflective optical systems based on physics-informed multi-task learning

The precision of optical mirror alignment is fundamental to ensuring optical system performance. However, traditional manual alignment is limited by its heavy reliance on expert experience, while existing data-driven approaches are constrained by their demand for massive high-quality datasets. To address these dual challenges, this study proposes a knowledge-driven adaptive alignment framework that integrates a fuzzy logic system with a physically constrained multi-task learning (MTL) neural network, enabling adaptive adjustment across varying mirrors. SHAP dependence analysis of simulation data is conducted to obtain a data-driven directional consistency prior, which is embedded into the MTL loss function as directional constraints. Subsequently, a fuzzy logic system is developed based on expert knowledge and its step-size outputs are dynamically modulated by the step size scaling factor learnt by MTL model. For the source domain mirror, the proposed method reduces the alignment iterations by 42.8% and decreases final peak-to-valley (PV) and root-mean-square (RMS) by 4.96% and 0.5% respectively, compared to a conventional fuzzy controller. Crucially, when applied towhat is preceived to be a new mirror model, the proposed method reduces the iterations by 55.6% and decreases PV and RMS by 6.9% and 13.7% respectively, compared with manual adjustment. This approach provides a robust, cross-model generalizable foundation for highly efficient automated optical assembly.