Mechanics-audited stateful operators for path-dependent finite-element digital twins

Path-dependent finite-element digital twins require predictions of both displacement and hidden plastic state. Accurate fields alone do not establish correct state evolution or mechanical equilibrium. We develop a mechanics-audited graph operator with coupled nodal and quadrature-point recurrent states. The operator predicts the full plane-strain plastic tensor and accumulated equivalent plastic strain, while assembled residuals, native energy, and local constitutive discrepancies regularize training. All core models share an elastic reference. A separate linear snapshot anchor retains the field-versus-mechanics comparison. The benchmark contains 240 independent material and loading trajectories on three complex quadratic meshes, with plastic activation and small-deformation range checks for every reference. Training, validation, and test sets are disjoint; 60 model runs use 50 epochs and five seeds. The 96-trajectory test set includes an unseen non-proportional path family. Against the architecture-matched data-only model, the audited operator changes mean active-state-increment error from 0.651 to 0.702 and residual ratio from 4.102 to 2.142. The joint residual and field-error benefit is not established. The residual-improvement direction persists under classical constitutive reintegration. The main model’s mean displacement relative error is 0.041; paired trajectory intervals and optimization variability are reported separately. Non-recurrent full-prefix and transformer-based constitutive baselines expose the remaining tradeoffs. The results support separate assessment of observable fields, plastic memory, and equilibrium within a qualified numerical domain; they do not establish universal model superiority or autonomous deployment reliability.