Vol.26, No.2, 2026, pp. 233–243
https://doi.org/10.69644/ivk-2026-02-0233 

Determining pile bearing capacity and settlement using analytical procedures commonly applied in engineering practice requires validation through field testing, either by static or dynamic methods. However, test results frequently exhibit significant deviations from calculated values. Since pile testing is both costly and time-consuming, this study investigates the application of eleven machine-learning algorithms to predict pile bearing capacity and settlement based on soil data at pile locations. The soil dataset is divided into laboratory and field data, forming four prediction approaches. The applied machine-learning algorithms include linear regression models, tree-based models, distance- and kernel-based models, as well as ensemble and boosting models. The database contains information on 170 piles tested both by static and dynamic methods within the Belgrade area. Algorithm performance is evaluated using R² values, with the CatBoost algorithm achieving the highest accuracy for both capacity and settlement prediction. For the best-performing model, the most influential parameters are identified for all four prediction approaches. The cross-sectional area of the pile is found to be the dominant factor in bearing-capacity prediction, while the applied load has the greatest influence on settlement prediction. Finally, conventional analytical methods for bearing capacity and set¬tlement are performed for all piles in the database. The results obtained using analytical calculation methods often exhibit significant deviations from those derived from pile load testing. By incorporating results from SLT and DLT and linking them with field and laboratory data, machine learning provides more advanced solutions compared to conventional analytical methods.

Keywords: • pile bearing capacity • pile settlement • pile testing • machine learning • geotechnical data