Thermodynamic-constrained neural modeling for path-dependent geomaterials via dynamic skeleton states

Addressing the dual challenges of high-fidelity data dependency and unphysical generalization in data-driven constitutive modeling, this study introduces a Reverse Thermodynamics-Informed Neural Network (Re-TINN). The framework establishes mappings between internal variables and both translational and dissipative stress to predict stress-strain responses, while incorporating dynamic origin position information as internal variables to characterize the evolution of soil fabric. By retroactively deriving Helmholtz free energy and dissipation rates, Re-TINN enhances thermodynamic constraints, thereby significantly improving predictive performance in scenarios lacking experimental free energy or dissipation rate data. The model is extended to drained and undrained conditions, with validation against synthetic and experimental datasets demonstrating: (1) Superior accuracy compared to conventional data-driven models for small-sample and noisy-data scenarios; (2) Enhanced capability to capture complex physics for path-dependent loading through the integration of dynamic reference states, as evidenced by variable importance analysis and accuracy quantification.