A physics-based correction surrogate complementing CUR226 for geosynthetic-reinforced pile-supported embankments

Abstract

Geosynthetic-reinforced pile-supported embankments (GRPSEs) are widely used for embankments over soft ground, but their response depends on complex coupled 3D mechanisms. Design methods such as CUR226 are widely used because they are fast and transparent, yet they rely on simplifying assumptions that can lead to differences relative to detailed numerical simulations and, for some design choices, conservative estimates of serviceability demand. Full 3D finite element (FE) analyses can better capture system response, but they are time-consuming for routine design screening and wide parametric exploration. This study proposes a practical, correction-based surrogate that retains CUR226 as a physics baseline and trains a simple neural network to learn the difference to a high-fidelity 3D FE database. A full- scale ABAQUS model, inspired by a well-documented field case, was used to generate 600 design cases spanning feasible ranges of key soil and geometrical input parameters. The model predicts log-transformed correction factors for differential settlement and maximum mobilized geosynthetic tensile force. Results show that the correction narrows the difference between CUR226 predictions and the FE responses. Level- averaged plots based on 10,000 Monte Carlo samples and Spearman’s rank correlations confirm that the qualitative trends with respect to the governing inputs are preserved, while response levels shift toward FE- consistent values. The surrogate enables rapid screening across a broad range of feasible configurations and provides a foundation for extending correction-based surrogates with richer datasets and improved learning models.