Machine learning-based prediction and interpretability of pull-out capacity of granular anchor piles in clay

This research uses machine learning (ML) techniques — ANN, MLR, RF, and SVR — to predict GAP’s pull-out capacity based on key parameters such as pile diameter, relative density, liquid limit, plastic limit, plasticity index, specific gravity, and [Formula: see text] ratio. A dataset from previous experimental studies is used to train and test these models, with their performance assessed using statistical metrics like [Formula: see text], NSE, PBIAS, MSE, and RMSE. [Formula: see text]-fold cross-validation was employed to evaluate the efficacy of the models. Visual comparison tools, such as regression plots, residual distribution, Taylor diagram and violin plots, were adopted for comprehensive model evaluation. Feature importance and the sensitivity of each input parameter on pull-out capacity were identified with RF and ANN, while SHAP plots help interpret the ANN predictions. The workflow uses python environment to implement training and validating the model. Artificial Neural Network (ANN) model with a ReLU activation function was identified as the best-performing method among all tested algorithms. Additionally, ML models are compared with two empirical equations for predicting pull-out capacity. Empirical equations can accurately predict the smaller pull-out value, but for larger values, these equations have not been able to converge. The SVR model performed worse than the empirical equations. However, ML models — except for SVR — exhibited high [Formula: see text] values, indicating greater accuracy. Feature Importance and SHAP analysis results show pile diameter as the most influential factor, followed by [Formula: see text] and soil consistency indices. Pile with optimum diameter with [Formula: see text] ratio[Formula: see text]8–10 and high density granular material ([Formula: see text]%) with moderate plastic clay yields the most favorable pull-out resistance, providing a guide for design and optimization of GAP in field conditions.