Interpretable machine learning for predicting pile capacity ratio: a case study of concrete piles in Iraq

Abstract The accurate determination of the ratio between design capacity and the measured capacity of piles is crucial for designing safe and cost-effective foundations. However, conventional methods for pile design rely on empirical equations which are unable to consider the complexity of soil-pile interactions due to lack of information about pile conditions and loading history. To overcome this issue, this study proposes a novel hybrid machine learning model named Tree-structured Parzen Estimator based Extreme Gradient Boosting (TPE-XGB) to estimate the effect of various pile and soil-related parameters including pile type, diameter, tip depth allowable pile on the ratio between design capacity from soil investigation and measured capacity from testing. For this purpose, 69 full-scale pile load tests were conducted, and GIS-based mapping was conducted to analyze spatial data. The TPE-XGB model was trained on the experimental data, and the results demonstrated the high efficacy of TPE-XGB model in predicting the output with 95% accuracy and minimal error (RMSE = 0.027). In addition, to interpret the findings of black-box TPE-XGB algorithm, Shapely Additive Analysis (SHAP) and Individual Conditional Expectation (ICE) analysis were used to identify the most important features and to explore the nonlinear relationships between input features and model output respectively. Finally, to facilitate the practical implementation of this study’s findings, a graphical user interface (GUI) was developed, allowing engineers to easily input site-specific parameters and obtain explainable model predictions. This data-driven approach offers a reliable tool for engineers to optimize pile design while ensuring transparency and reliability in decision-making.