Influence of feature selection on machine learning prediction of pile foundation – The role of soil-pile interaction knowledge and application to base resistance

The use of machine learning (ML) to predict and classify various aspects of pile foundation has rapidly gained paramount attention over the past decade, resulting in various sets of input features and algorithms that have been employed. This study classifies those input features into 2 groups, i.e., with and without understanding of soil-pile interaction, and examines their impacts on the development and performance of ML models. Six different ML algorithms covering from the most fundamental decision trees to advanced bagging and boosting techniques are adopted. An extensive database gathering 86 cases (1129 datapoints) of pile load tests in complex soft soil region is developed and employed to investigate the influence that 2 different sets of input features can have on predicting base resistance of piles. Through a comprehensive training-to-validation process, the results prove the importance of having soil-pile interaction features such as the load-displacement curves in optimizing the ML prediction of base resistance. On the other hand, the use of only soil and pile features for model inputs can help reduce the cost for pile tests, though it can lead to poorer prediction performance. Not only does the current study establish novel models based on machine learning to predict base resistance of piles, but it also employs quantitative feature analysis to gain insight into the load-transfer process, where the interaction between soil and pile develops downward from side to base resistances. The study proposes an adaptive training strategy that effectively improves model application to new contexts, reducing the need for extensive data collection and field survey costs, thereby enhancing the cost-effectiveness and scalability of ML deployments in geotechnical engineering.