Self-adaptive physics-informed parallel neural networks for creep damage identification

Identification of creep damage poses a significant challenge in structural engineering due to its nonlinear, time-dependent stress redistribution, especially under limited and noisy measurements. Building on our previous work on Physics-Informed Parallel Neural Networks (PIPNNs) and Neural Tangent Kernel (NTK)-based self-adaptive loss weighting, this paper extends the framework for the inverse identification of creep damage governed by continuum damage mechanics. The proposed self-adaptive PIPNN integrates creep strain and damage evolution laws directly into the physics-informed loss function. A domain decomposition strategy enables scalable modeling across complex subdomains, while NTK-based loss weighting dynamically balances competing loss terms and mitigates numerical stiffness. By minimizing the physics-informed loss, the method simultaneously estimates initial damage parameters and reconstructs full-field structural response. Validation on three examples, a two-bar system with creep rupture, a planar truss with localized damage, and a two-span Euler–Bernoulli beam with distributed degradation, demonstrates accurate damage identification and robust response prediction under limited and noisy data. These results highlight the framework’s potential for physics-constrained structural health monitoring and lifetime forecasting. • A self-adaptive PINN framework is developed for creep damage identification. • Creep and damage laws are embedded into the PINN architecture. • Initial damage is estimated from limited and noisy structural measurements. • Full-field response is reconstructed with high accuracy and robustness. • Demonstrated on bar, truss, and beam systems with creep damage evolution.