A Machine Learning-Based Method with Integrated Physics Knowledge for Predicting Bearing Capacity of Pile Foundations

The ultimate bearing capacity is one of the main parameters in pile foundation design. Typically, design load calculations can be confirmed by static or dynamic load tests. The static load test is often considered the most reliable method to obtain the ultimate bearing capacity; however, this method is often labor-intensive, time-consuming, and cost-ineffective. Thus, it is of practical interest to predict the ultimate bearing capacity of pile foundations efficiently and accurately in the design phase. This paper presents a machine learning-based framework with integrated physics knowledge for predicting the bearing capacity of pile foundations. A database with 200 static load tests of pile foundations from the literature was used. Several commonly used machine learning (ML) algorithms were trained and tested, including ridge regression, support vector machine, random forest, and gradient boosting machine. As ML models learn functional relationships based on data, trained models with limited data often have unexpected behavior when predicting out-of-domain samples, reducing the reliability of ML models in engineering practice. In this study, prior knowledge in geotechnical field was integrated into these ML models, where shape constraints and additional features were applied to the input space during the model training stage. The results show that the developed ML models have better performance in predicting the bearing capacity than traditional semi-empirical method, and the models integrated with physics knowledge have improved performance.