J-PINN: A domain-decomposed physics-informed neural network framework for kinematic field reconstruction in fracture mechanics

Physics-Informed Neural Networks (PINNs) solve inverse problems, yet standard formulations fail to resolve the discontinuous fields inherent to fracture mechanics. This study introduces a novel, meshless Multi-Region PINN framework for 2D cracked elastic bodies. By decomposing the domain via strict kinematic constraints, the method restores the validity of the Navier–Cauchy equations within each partition. Validated against FEM, the framework accurately captures stress singularities using sparse boundary data, achieving relative errors below 1% for displacements and 2–4% for strain and stresses. Furthermore, a thermodynamic error-compensation strategy is proposed to correct systematic -integral drift, yielding path-independent strain energy release rate (SERR) estimates. Demonstrating utility for material characterization, the method is applied to experimental Digital Image Correlation (DIC) data from a Glass Fiber Reinforced Polymer (GFRP) Double Cantilever Beam. Here, the PINN acts as a physics-informed reconstructor, recovering smooth singular fields from noisy inputs via a robust Huber loss and a physics-driven iterative outlier removal strategy. The calculated SERR ( ) shows strong agreement with analytical Modified Beam Theory (MBT), establishing a robust, assumption-free characterization methodology for interlaminar fracture toughness in fiber-reinforced composites and enabling data-driven extraction of cohesive zone parameters as faithful inputs for composite structural design.