Inverse identification of multiple constitutive parameters of composites utilizing physics-informed neural networks

Composite materials are widely used in aerospace, automotive, and renewable energy applications owing to their superior mechanical properties. However, accurately characterizing their constitutive parameters, especially the shear nonlinearity in thick-section composites (TSCs), remains a challenging task. This study proposes a Physics-Informed Neural Networks (PINN) framework that, for the first time, enables simultaneous inverse identification of multiple constitutive parameters through one single experiment. The framework constructs a physics-augmented multi-objective loss function that explicitly incorporates equilibrium equations, nonlinear constitutive relations, and boundary conditions as physical constraints, while integrating full-field strain as data-driven terms. This design facilitates synergistic optimization of strain field reconstruction, parameter identification, and stress field extraction through physics-guided joint training. To enhance stability and generalization, a transfer-learning-inspired strategy is employed by fine-tuning a pre-trained forward PINN. Validation via simulated three-point bending experiments demonstrates exceptional accuracy, with maximum relative error below 0.41%, and robustness against variations in initial guesses, network architectures, and collocation sampling densities. Benchmarking against the conventional Finite-Element-Model-Updating (FEMU) approach highlights the superior performance of the proposed framework: it achieves a significant improvement in inversion accuracy, reducing the mean relative error from 6.766% (FEMU) to 0.959%, while attaining a 5.75-fold speed-up in convergence. Furthermore, the study underscores the importance of a carefully designed loss formulation that effectively couples data-driven supervision with physics-informed constraints. The presented framework is not only applicable to thick-section composites but also readily extendable to other anisotropic materials and complex nonlinear constitutive relations, thereby offering a reliable and efficient computational tool for advanced material characterization.