Abstract: In recent years, efforts have been made to apply the Artificial Intelligence for Science (AI4S) paradigm in various fields of hydraulic and hydropower engineering, e.g. the data-driven techniques used in the constitutive modeling of engineering materials. However, data-driven constitutive models often suffer from limited generalizability and robustness; most of the previous studies remained confined to simple numerical examples, leaving applicability to complex engineering problems in need of further verification. This study adopts a Generalized Plasticity Model-Physics-encoded Neural Network (GPM-PeNN), developed by our team, to simulate the stress and deformation of a rockfill dam. This model is trained using a synthetic dataset of rockfill materials from the Lawa high concrete-faced rockfill dam, and it is embedded into the general-purpose finite element code ABAQUS via a user-defined material module (UMAT). It is used to simulate the stress and deformation responses during dam-filling. Compared with finite element analyses based on traditional constitutive models, our simulations－based on the physics-encoded neural network constitutive model－align with general mechanical behaviors, and exhibit high accuracy and good convergence, thereby validating the feasibility of applying data-driven constitutive models in practical engineering applications.

Key words: data-driven, constitutive model, physics-encoded, rockfill dam, stress-deformation analysis