Deep Learning Surrogate for Undrained Cyclic Response of Sands: Stability and Generalization

Predicting the history-dependent cyclic response of granular soils remains a central challenge for data-driven methods, which frequently rely on synthetic training data or exhibit instability during the post-liquefaction regime. This study presents a Long Short-Term Memory (LSTM) framework trained exclusively on raw experimental databases from cyclic Direct Simple Shear (DSS) and Triaxial (TXC) tests on Ottawa and Karlsruhe sands. To overcome the computational and physical constraints of laboratory data, the training pipeline implements systematic data thinning and physics-guided loss weighting, ensuring the precise capture of early-stage pore-pressure generation. Validated through rigorous Leave-One-Test-Out cross-validation, the model robustly predicts unseen stress–strain–pore pressure trajectories, including the broad, highly dissipative hysteresis loops characteristic of post-liquefaction flow without evidence of cumulative drift. Comparative benchmarks demonstrate that the restructured LSTM performs comparably to the PM4Sand constitutive model in reproducing stiffness degradation under different cyclic stress ratios. Furthermore, error diagnostics identify fabric anisotropy as the primary governor of prediction limits in dense sands. This research establishes a stable constitutive surrogate capable of generalising across diverse material states within each loading mode.