A methodology to automate precision optical alignment using sensitivity-aware optimization

In optical and laser systems, accurate optical alignment is critical, as sub-micron positioning errors can lead to beam degradation, reduced throughput, and unstable system behaviour. Therefore, in this study, we introduce a generalised combined machine learning and optimisation approach that (a) models how alignment adjustments influence optical propagation and (b) computes real-time correction commands. The model is trained on simulated alignment data and then adapted to variations in the optical configuration using transfer learning. This approach reduces the need for full retraining after any configuration changes. A sensitivity analysis identifies the most influential actuators, enabling the system to reduce correction dimensionality and improve both convergence speed and stability. We present an example combining supervised learning with a ray-based physical simulation of a laser resonator with actuator updates computed using a constrained nonlinear optimisation routine that maximizes laser pulse energy. The proposed approach achieves a sub-50 ms control-cycle execution in our implementation, which allows for rapid correction. The modular framework enables real-time control and is compatible with standard robotic or mechatronic stages and custom actuator configurations. It also demonstrates generalisability by providing rapid and reliable alignment across automated testing environments, precision optics manufacturing, and research setups.