Physics-based neural networks for coupled thermo-hydro-mechanical-chemical analysis of FEBEX bentonite

Long-term geotechnical studies are often constrained by coupled multiphysics modelling ability and sparse experimental datasets due to the extended time scales and harsh environmental conditions involved. To address those limitations, this study presents an application of Physics-Based Neural Networks (PBNNs) for modeling the complex coupled thermo-hydro-mechanical-chemical (THMC) processes in bentonite barrier systems used in nuclear waste repositories. The methodology integrates sparse experimental data points for calibration with physics-compliant synthetic dataset generated from well-established physical models, including Arrhenius kinetics for chemical reactions, diffusion equations for mass transport, thermal evolution models, consolidation theory for mechanical behavior, and chemical equilibrium principles. This approach ensures that the synthetic data maintain physical constraints and consistency while providing sufficient training samples for neural network optimization. Seven neural network models were trained using this synthetic dataset to predict the long-term evolution of temperature, water content, porosity, chemical concentrations, and mechanical properties. The trained models achieved excellent performance against sparse experimental data points with R² scores exceeding 0.96 for all variables, demonstrating minimal overfitting (generalization gap < 0.005) despite the original data scarcity. Key findings include: (1) accurate prediction of exponential temperature evolution consistent with thermal physics, (2) successful modeling of water content decay following Arrhenius kinetics, (3) realistic simulation of porosity evolution through coupled hydration, consolidation, and pore refinement processes, and (4) robust prediction of chemical species transport and mechanical stress development. This PBNN approach provides a practical solution to data scarcity in geotechnical engineering, offering significant computational advantages (180–720× speedup) over traditional numerical methods while maintaining physical consistency. The methodology enables automatic learning of complex coupling relationships without requiring explicit physics loss terms during training. This work demonstrates a viable computational framework for long-term safety assessment of nuclear waste repositories and highlights the potential of PBNN in sparse data modelling of engineering applications.