Abstract: The material constitutive model is the basis of numerical analysis in geotechnical mechanics. Phenomenological constitutive models, represented by elastoplastic models, are the mainstream of geotechnical constitutive research. In recent years, data-driven methods such as neural networks have increasingly become an important means of constitutive modeling owing to their strong nonlinear fitting capability. However, conventional data-driven models, lacking prior physical knowledge, are prone to poor generalization and physical inconsistency. This paper proposes an improved neural network-based constitutive modeling method that explicitly incorporates mechanical mechanisms including thermodynamic consistency and plastic dissipation stability within the elastoplastic mechanics framework. Taking the von Mises model and the modified Cam-clay model as examples, diverse stress paths are generated by Gaussian process to construct training sets, and the performances of purely data-driven and mechanics-constrained models are systematically compared. The results show that, with the introduction of mechanical constraints, the prediction accuracy and stability of the model under unseen paths are significantly improved. The engineering applicability and generalization potential of the proposed method are verified by finite element integration tests.

Key words: physics-constraints, neural networks, data-driven modeling, generalization capability, finite element method