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

Area of Science:

• Geotechnical Engineering
• Machine Learning Applications
• Foundation Design

Background:

• Micro-piled raft foundations are crucial in clay soils.
• Understanding their load-settlement behavior is complex due to nonlinear soil-structure interaction.
• Accurate prediction models are needed for efficient geotechnical design.

Purpose of the Study:

• To evaluate machine learning algorithms for predicting the load-settlement behavior of micro-piled raft foundations in clay.
• To identify the most accurate and reliable predictive model.
• To assess the impact of geometric and soil properties on foundation performance.

Main Methods:

• Utilized a dataset of 480 experimental records from lab and field tests.
• Evaluated six supervised machine learning algorithms: GPR, XGBoost, GBM, RF, KNN, and SVR.
• Optimized models using Bayesian optimization and 5-fold cross-validation; validated with Monte Carlo simulations.

Main Results:

• Gaussian Process Regression (GPR) exhibited superior accuracy and generalization capabilities.
• GPR effectively captured the complex nonlinear soil-structure interactions.
• Probabilistic analysis confirmed the robustness and reliability of GPR predictions.

Conclusions:

• GPR is a highly promising tool for practical geotechnical design of micro-piled raft foundations.
• The model demonstrates reliability through close agreement with experimental data and narrow confidence intervals.
• Future work should focus on dataset expansion and advanced tuning for broader applicability.