Optimizing reinforced pile-supported embankments using surrogate-assisted FE modeling

This study presents a proof-of-concept surrogate-assisted finite-element (FE) framework for multi-objective design optimisation of geosynthetic-reinforced pile-supported embankments (GRPSE), focusing on internal stability. A multilayer perceptron neural network was trained using data generated from an automated 2.5D full-scale FE model developed in ABAQUS through Python scripting. The design variables included pile spacing, pile diameter, gravel-platform friction angle, and embankment height. The trained surrogate was integrated into a multi-objective optimisation framework to identify the cheapest, safest, and most balanced GRPSE designs using Pareto fronts in the cost-safety space. Optimised solutions were compared with those from the established CUR226 analytical framework. Results show that, for the cheapest and safest designs, the surrogate model provides cost and safety performance comparable to CUR226 while maintaining computational efficiency during optimisation. By capturing complex soil–structure interactions, it also identifies alternative balanced combinations of design variables. For the optimal design, the surrogate model achieves a higher factor of safety at comparatively lower cost. Although CUR226 remains effective, particularly when fill material properties govern the design, the surrogate framework offers added value for pile-soil interaction-informed optimisation. Future work should integrate CUR226 as guiding physics within neural network models trained on modest datasets for practical design.