Path-aware data-driven computational mechanics for elastoplastic bedded rock

Bedded rock exhibits pressure-dependent yielding, strong sensitivity to stress path, and pronounced anisotropy associated with bedding orientation, making it difficult to describe with conventional phenomenological plasticity models. Data-Driven Computational Mechanics (DDCM) provides a model-free alternative by replacing explicit constitutive laws with nearest-neighbor retrieval from a material-state database. However, two fundamental issues constrain its application to elastoplastic geomaterials. First, the conventional fixed-scale distance metric systematically favors elastic database entries in the post-yield regime, where accumulated plastic strains may exceed the elastic reference scale by one to two orders of magnitude. Second, because plastic response depends on loading history, stress proximity alone cannot reliably distinguish states with different paths, leading to physically inconsistent nearest-neighbor matches near the yield surface. We develop a unified path-aware DDCM framework to address these two issues without introducing an explicit yield criterion or flow rule. We adapt the strain reference locally at each node according to its current database assignment, removing the bias toward elastic entries in plastic regions. To account for loading-path effects, we include confining pressure as a surrogate history variable in both the database entries and query vectors. Before each nearest-neighbor search, a Mixture Density Network (MDN) conditioned on partial observations completes the query in a branch-aware manner, resolving the bimodal structure of the elastoplastic material manifold. The physical projection is formulated in terms of displacement, eliminating the kinematic compatibility constraint and improving convergence along strongly nonlinear loading paths. The framework is assessed against Mohr–Coulomb analytical solutions for a thick-walled cylindrical cavity at five bedding angles ( \(\:\beta\:\) = 0°, 30°, 45°, 60°, 90°). Stress-subspace retrieval recovers the discontinuous strain jump at the elastic-plastic interface directly without specifying an explicit jump condition. At \(\:\beta\:\) = 30°, increasing the training pressure range by 20% reduces the circumferential stress RMSE by 67%, indicating that database coverage, not database density, is the major limitation on solution accuracy. Increasing the database size from 500 to 48,000 state pairs shows little change in solution accuracy, while the computational cost increases only sublinearly with database size. These results, validated against analytical elastoplastic solutions across five bedding orientations, show the framework performs consistently when using separate orientation-specific databases with the same optimization algorithm. While this validation establishes algorithmic correctness in analytically tractable configurations, extending to experimental material data and cross-validating with alternative constitutive models remain important future steps to assess practical robustness under measurement noise and model-form uncertainty.