Physics-Informed Deep Learning for Automated Defect Localization and Material Characterization in Ultrasonic Non-Destructive Testing

Abstract

Ultrasonic non-destructive testing remains a cornerstone of structural integrity assurance across aerospace, energy, transportation, and civil infrastructure. Conventional interpretation of ultrasonic signals depends heavily on expert judgment, physics-based inversion, and controlled inspection protocols. Deep learning has created new possibilities for automated defect localization and material characterization, but purely data-driven approaches often suffer from limited labeled data, sensor variability, and physically inconsistent predictions. Physics-informed deep learning integrates elastic wave propagation principles, scattering constraints, and material behavior into neural network training and architecture design. This paper presents a system-level examination of physics-informed deep learning for ultrasonic non-destructive testing. It analyzes data acquisition infrastructure, model architecture, simulation-assisted training, deployment constraints, robustness, uncertainty quantification, governance, and sustainability. The discussion emphasizes structural trade-offs between physical fidelity and computational cost, centralized training and edge deployment, and automation and certification. By treating the inspection pipeline as a socio-technical system, the paper argues that industrial adoption requires modular design, standardized data provenance, continuous recalibration, and transparent decision support. Cross-domain comparisons with medical imaging and structural health monitoring further illustrate how governance mechanisms and uncertainty management can support trustworthy automation. The analysis concludes that physics-informed deep learning can improve defect localization and material characterization only when embedded within an institutional framework that supports validation, lifecycle maintenance, and regulatory alignment.