Physics-informed identification of stress fields and thermo-viscoplastic model parameters for metals from full-field data under impact loading

Metals and alloys exhibit highly nonlinear thermo-mechanical behavior under impact loading, strongly influenced by strain rate and adiabatic heating. Accurate calibration of thermo-viscoplastic constitutive parameters traditionally requires extensive stress–strain measurements across a wide range of loading rates and temperatures. Recent advances in full-field measurement techniques enable rich kinematic and temperature data acquisition in dynamically loaded structures, and machine learning techniques have shown promise for stress and parameter identification. However, most existing efforts remain limited to quasi-static loading scenarios. In this study, we propose a family of physics-informed neural network (PINN) frameworks—T-PINN, ε-PINN, and T-ε-PINN—named after the types of dynamic full-field data they employ. These frameworks reconstruct stress fields and identify thermo-viscoplastic parameters in metals by embedding the governing equations of thermo-elasto-viscoplasticity. Specifically, T-PINN combines temperature field data with linear momentum balance, energy conservation, and constitutive relations to infer stress distributions. The ε-PINN incorporates strain field data to simultaneously reconstruct stress fields and identify thermo-viscoplastic constitutive parameters, accounting for strain hardening, strain-rate sensitivity, and thermal softening effects under dynamic loading. The T-ε-PINN integrates both temperature and strain-field measurements for enhanced data efficiency. Through comprehensive numerical examples, we demonstrate the effectiveness of these methods in accurately reconstructing stress fields and predicting material responses under complex dynamic loading conditions, even in the presence of noisy data. These PINN-based approaches provide a data-efficient, physically well-defined alternative to traditional calibration methods for dynamic constitutive modeling of metallic materials.