Physics-informed neural networks for phase-based material defect identification

The accurate identification of material defects is critical for ensuring structural integrity and performance. Traditional computational methods often struggle to balance efficiency and physical fidelity in complex material systems. This paper presents a novel approach integrating physics-informed neural networks (PINNs) with the phase field method to address these challenges. Our approach leverages a phase field variable to delineate intact regions from voids, while a stress degradation model modifies mechanical responses at defect sites. Neural networks serve as surrogate forward solvers to predict displacement and stress fields, enabling rapid simulations. To ensure compatibility with physical laws, the framework embeds governing equations into the training loss function. Additionally, a data-driven term minimizes discrepancies between simulated and experimentally measured strain fields, enhancing defect localization precision. Numerical experiments validate the framework’s robustness across diverse configurations, including circular, elliptical, irregular, and multiple voids, as well as material behaviors, extending from linear elastic to hyperelastic models. The results demonstrate superior accuracy in identifying void geometry, size, and spatial distribution compared to conventional methods. The proposed approach’s adaptability to complex geometries and material nonlinearities highlights its broad applicability in aerospace, automotive, and biomedical industries.