Settlement-informed boundary-adaptive excess pore water pressure prediction during vacuum preloading

Abstract

Predicting excess pore water pressure (EPWP) fields during vacuum preloading remains difficult because EPWP sensors are sparse and the time-varying vacuum boundary is rarely recorded continuously. This study frames this problem as settlement-informed boundary adaptation rather than direct full-field validation. A site-constrained radial-vertical finite-difference method (FDM) operator is first calibrated to the Shanghai Disneyland case and used to generate baseline EPWP and settlement responses. The monitored-FDM settlement discrepancy is then decomposed to infer bounded equivalent vacuum-boundary disturbances, which are back-substituted into the FDM operator to generate physically propagated non-ideal EPWP fields. A boundary-adaptive PGRL-ConvLSTM combines FDM one-step propagation, neural residual correction, GRU-based boundary-state estimation, soft top-boundary writeback and block-wise rollout. The FDM baseline reproduced the monitored settlement at 60.9 d and showed representative EPWP agreement at three monitored depths, but these checks are treated as calibration-consistency and representative constraints rather than independent full-field validation. Under ideal FDM boundary conditions, the terminal RMSE decreased to 0.239 kPa. Under non-ideal blind-boundary forecasting, the overall RMSE decreased from 1.441 to 1.064 kPa relative to the hard-boundary variant. Ablation and frozen shifted-case tests further indicate that boundary-state adaptation mainly improves disturbance-related boundary errors. The evidence is limited to one field case and FDM-generated full-field labels; future validation requires denser EPWP and vacuum-boundary monitoring.