Data-driven prediction of micro-piled raft load-settlement using machine learning and Monte Carlo simulation.

This study investigates the load-settlement behavior of micro-piled raft foundations in clay, focusing on key factors such as raft and micro-pile geometry and critical soil properties. A comprehensive dataset comprising 480 experimental records. sourced from both small-scale laboratory and large-scale field tests. was used to evaluate the predictive capabilities of six supervised machine learning algorithms: Gaussian process regression (GPR), extreme gradient boosting (XGBoost), gradient boosting machine (GBM), random forest (RF), K-nearest neighbors (KNN), and support vector regression (SVR). Each model was optimized using Bayesian optimization with 5-fold cross-validation to ensure robust performance. Model evaluation was conducted using statistical metrics, visual diagnostics (predicted-versus-actual plots), Regression error characteristics curves, score analysis, and hyperparameter tuning. Among the tested models, GPR demonstrated superior accuracy and generalization, effectively capturing the nonlinear soil-structure interaction typical of micro-piled raft foundations. Probabilistic analysis using Monte Carlo simulations, incorporating realistic variability in geometric and geotechnical input parameters, further validated the model's robustness. The close agreement between predicted and experimental load-settlement responses, along with consistently narrow 95% confidence intervals, confirms the reliability of GPR. These findings position GPR as a highly promising approach for practical geotechnical design applications. Future research should focus on expanding dataset diversity, improving parameter independence, and applying advanced tuning techniques to further enhance model reliability and applicability to full-scale foundation systems.