Thermodynamically consistent modeling of granular soils using physics-informed neural networks.

In recent years, data-driven approaches have gained considerable momentum in the scientific and engineering communities, owing to their capacity to extract complex patterns from high-dimensional data. Despite their potential, these approaches often require extensive high-quality datasets, may exhibit limited extrapolation capability beyond the training domain, and lack a rigorous foundation grounded in physical and thermodynamic principles. To overcome these limitations, physics-informed neural networks have been introduced, embedding governing equations directly into the learning process. Building upon this paradigm, this study presents a novel thermodynamically consistent constitutive model for granular soils, developed within the framework of geotechnically- and physics-informed neural networks (GINN). The model integrates physical laws with data-driven learning via a composite loss function. These include: (i) strictly non-negative material dissipation rate to ensure thermodynamic admissibility, (ii) an admissible range for the predicted stress state, and (iii) bounds on the predicted void ratio. The material dissipation rate is calculated using the total work input and a free energy potential expressed in terms of stress invariants. The model is validated against monotonic drained triaxial test data for specimens prepared with diverse initial void ratios and stress states. The model accurately simulates both the shear strength and dilative response of granular soil samples. Its predictive performance is further benchmarked against two widely adopted constitutive models from the literature, demonstrating comparable accuracy while maintaining consistency with thermodynamic laws.