Field-aware physics-informed transformer for robust computer-aided alignment of off-axis TMA system with mirror...

Area of Science:

• Optical Engineering
• Machine Learning for Optics
• Space Optics

Background:

• Off-axis reflective optical systems are crucial for space exploration and high-resolution imaging due to their compact, unobscured design.
• System asymmetry causes severe aberration coupling, challenging traditional alignment methods and deep learning approaches with unstable convergence or failure.
• Existing methods struggle with the complex interplay between misalignments and surface errors in these systems.

Purpose of the Study:

• To develop a robust and accurate automated alignment framework for off-axis reflective optical systems.
• To address the limitations of current methods in handling aberration coupling and achieving stable convergence.
• To establish a physically interpretable and effective solution for complex optical system alignment.

Main Methods:

• Proposed a field-aware physics-informed transformer (FA-PIT) model.
• Reformulated Zernike coefficients as structured field tokens to preserve spatial topology.
• Employed a multi-head self-attention mechanism for learning cross-field aberration coupling and a physics-informed composite loss function for robust compensation.

Main Results:

• FA-PIT achieved high success rates (99%, 96%, 69%) under various error conditions, significantly outperforming five baseline methods.
• Demonstrated superior performance in out-of-distribution conditions, retaining high success rates (73%-90%) where others failed.
• Showcased robustness against measurement noise and cross-system transferability with a 93% success rate on an independent architecture.

Conclusions:

• FA-PIT provides an effective, physically interpretable, and robust framework for automated alignment of off-axis reflective optical systems.
• Key components like field tokenization, self-attention, and physics-informed loss are essential for FA-PIT's performance gains.
• The model offers a significant advancement over traditional and current deep learning alignment techniques.