This study is on the application of ensemble machine learning techniques in stonecolumn bearing capacity prediction, integrating experimental and field test dataacross various soil conditions to address the limitations of traditional methods. Thispaper developed and evaluated several predictive models, including linear regression,gradient boosting, support vector regression, K-nearest neighbors, and a newlyproposed ensemble model. The results revealed that traditional linear regressionproduced moderate accuracy (R2 = 0.848, RMSE = 19,220.12, MAE = 14,449.63),while support vector regression substantially underperformed (R2 = 0.169,RMSE = 53,378.71, MAE = 20497.13), underscoring the challenges posed by dataheterogeneity and nonlinearity. In contrast, ensemble approaches, exemplified bygradient boosting (R2 = 0.986, RMSE = 5806.08, MAE = 1697.49) and K-nearestneighbors (R2 = 0.987, RMSE = 5593.26, MAE = 1668.71), demonstrated remarkableimprovements in accuracy. K-nearest neighbors emerged as the clear front-runneramong the models tested, achieving the highest explained variance alongside thelowest error metrics. Its predictions were exceptionally stable; over 90% of residualsfell within ± 10% of actual capacities across all load ranges, and both trainingand validation R2converged above 0.98 as the dataset expanded, indicating minimaloverfitting. This combination of peak accuracy, consistency under diverse soil conditions,and strong generalization made K-nearest neighbors the preferred choice forrapid, reliable estimation of stone-column bearing capacity in practical geotechnicaldesign.
