Thermodynamics‐Informed Neural Identification of Path‐Dependent Interface Laws From Boundary‐Only Observations

Progressive interface degradation governs the long‐term load‐bearing capacity of fully grouted rock anchors, yet field pullout tests yield only boundary load and displacement. Internal stress and damage fields remain experimentally inaccessible. This article presents TIINN (thermodynamics‐informed interface neural network), a framework that provides a thermodynamically constrained estimation of path‐dependent interface constitutive laws from boundary‐only cyclic pullout observations. TIINN couples a Bregman energy network, a bounded memory damage operator (), and a field distillation network with a differentiable finite‐difference solver. Equilibrium is satisfied to numerical tolerance via Newton iteration, and gradients propagate to constitutive parameters through implicit differentiation, eliminating the PDE/BC loss‐weight tuning inherent to penalty‐based physics‐informed methods. Synthetic benchmarks confirm high‐fidelity constitutive recovery with boundary and damage‐field correlation . Ablation demonstrates that thermodynamic architectural constraints reduce fit variability by two orders of magnitude relative to unconstrained networks. Application to field rock‐anchor cyclic pullout data improves the in‐sample by 0.44 over the no‐damage baseline. However, leave‐one‐cycle‐out prediction yields , revealing that six‐cycle boundary‐only protocols provide insufficient information‐theoretic constraint for reliable forward extrapolation of damage evolution. The framework establishes a transferable methodology for indirect constitutive identification in embedded interface systems where internal measurements are unavailable.