Physics-informed computational imaging for newton’s rings: a benchmark and observability analysis

Abstract

Physical parameter estimation from interferometric fringe patterns remains a challenging inverse problem due to the complex coupling of wave optics and sensor-level degradations. Traditional analytical methods often rely on idealized assumptions of radial symmetry and monochromaticity, limiting their robustness under severe spectral decoherence, geometric misalignment, and low-observability ( Veff<0.7) conditions. To connect theoretical interference models with practical sensor observations, we propose a physics-informed computational imaging framework and a comprehensive benchmarking protocol for Newton’s rings analysis. Our approach integrates a differentiable wave-optics forward model with a sensor-aware degradation pipeline, enabling end-to-end gradient-based inference of physical parameters. Crucially, we introduce an effective visibility metric ( Veff) to quantify the practical observability threshold for recoverability of physical parameters under broadband illumination. Our analysis reveals a performance transition at Veff≈0.7, below which deterministic parameter estimation becomes ill-posed owing to signal decoherence, defining the lower bound on parameter recoverability under broadband illumination. Extensive evaluations demonstrate that embedding physical inductive biases significantly enhances the robustness of deep learning models, particularly in low-observability regimes where classical fringe analysis methods degrade catastrophically. We release the complete dataset and differentiable simulation engine to facilitate reproducible research in physics-AI co-design for interferometric imaging.