Transformer-based settlement prediction model of pile composite foundation under embankment loading

Pile composite foundation (PCF) is a common method for treating weak foundations, with settlement being the primary indicator in its design. However, accurately and quickly obtaining PCF settlement remains challenging. This study proposes a transformer-based PCF settlement prediction model under embankment loads (PCFFormer model), which enables efficient and accurate predictions of PCF settlement across various embankment environments and pile schemes. To establish and validate the data-driven PCFFormer model, this study also developed an automatic modeling and data processing program for PCF based on the Abaqus platform. Furthermore, a large-scale dataset of PCF finite element models under embankment loads was constructed and released, which is currently the largest publicly available dataset of PCF finite element models. Additionally, this study introduces a data augmentation method that fully utilizes the finite element analysis process data, significantly improving the efficiency of creating the PCF settlement dataset. By comparing the PCF settlement predicted by the PCFFormer model with the results of finite element analysis and the predictions of other machine learning methods, the accuracy and superiority of the PCFFormer model are demonstrated. The study further discusses the impact of missing individual parameters in cushion and soil layers on the prediction accuracy of the PCFFormer model.