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

Area of Science:

• Optical Engineering
• Artificial Intelligence
• Robotics

Background:

• Manual optical mirror alignment relies heavily on expert experience.
• Existing data-driven methods require extensive, high-quality datasets.
• Automated optical assembly demands efficient and generalizable alignment techniques.

Purpose of the Study:

• To develop a knowledge-driven adaptive alignment framework for optical mirrors.
• To overcome limitations of manual and purely data-driven alignment methods.
• To enable robust, cross-model generalizable automated optical assembly.

Main Methods:

• Integration of a fuzzy logic system with a physically constrained multi-task learning (MTL) neural network.
• Utilizing SHAP dependence analysis for data-driven directional consistency priors.
• Embedding directional constraints into the MTL loss function and dynamically modulating fuzzy logic step-size outputs.

Main Results:

• Reduced alignment iterations by 42.8% and improved peak-to-valley (PV) and root-mean-square (RMS) metrics compared to conventional fuzzy controllers.
• Achieved 55.6% fewer iterations and decreased PV and RMS by 6.9% and 13.7% respectively on a new mirror model compared to manual adjustment.
• Demonstrated significant efficiency gains and precision improvements in optical mirror alignment.

Conclusions:

• The proposed knowledge-driven adaptive alignment framework offers a robust solution for automated optical assembly.
• The integration of fuzzy logic and MTL provides efficient and generalizable mirror alignment capabilities.
• This approach significantly enhances precision and reduces iteration counts in optical system performance tuning.