Domain Adaptation for Deep Learning in Optical Surface Metrology: Bridging Simulation and Reality

Abstract

Deep learning models trained on simulation data achieve strong performance in optical surface metrology tasks, but often suffer significant performance degradation when deployed on real experimental measurements due to the distribution shift between simulated and real data. This domain gap—between the synthetic training domain and the physical measurement domain—represents one of the principal barriers to practical deployment of deep learning in precision optical metrology. This study proposes a domain adaptation framework for optical surface metrology that enables deep networks trained on simulation data to generalize effectively to real experimental measurements, building upon the deep learning methodologies established by Huang, Tang, Liu, and Huang (2026) in optical metrology and the 4D thermal imaging approach of Huang, Yang, and Zhu (2023). The framework employs an unsupervised domain adaptation strategy combining a physics-based appearance transfer module with an adversarial domain discriminator, enabling the network to learn domain-invariant feature representations without requiring labeled real-world data. The approach is validated across three representative tasks: thermal image reconstruction on non-flat surfaces, phase unwrapping in deflectometry, and surface defect detection. Simulation-to-reality transfer experiments demonstrate that the proposed framework reduces performance degradation at deployment by 73% on average compared to models deployed without domain adaptation, and achieves within 8% of the performance of models trained on fully labeled real experimental data. The framework provides a practical pathway for transferring deep learning solutions from simulation to production-line optical inspection systems.