The ultimate bearing capacity (UBC) of piles is a function of soil parameters, pile material and dimensions and the intricacies of soil pile interactions. Reliability design of pile foundation involves the use of a combination of conventional methods for the design and verification of pile capacity. Machine learning (ML) can employ historical data to develop predictive models for new or similar conditions. In this study, the K-nearest neighbours (KNN) and KStar algorithms, modelled in an open-source software WEKA, were deployed for the prediction of UBC of piles. The dataset consists of 100 instances of both steel and concrete piles with cohesion, drained friction angle, effective soil weight, internal friction angle, flap, area of pile, length of pile and pile material as input features, and UBC as the output feature. The methodology includes data visualisation and variable importance analysis and model assessment. The optimal model yielded a superb performance with root-mean-square error, RMSE and correlation coefficient values of 1065.36 and 0.9986, respectively, both representing some improvements to the outcome of previous studies. Sensitivity analysis also revealed that the length of pile is the most important feature in the model, and this aligns with pre-modelling assumptions based on data visualisation. Overall, this research affirms the capability of WEKA, and the efficacy of its KNN and KStar algorithms for the prediction of the bearing capacity of piles. It is a reliable tool for piles UBC prediction that can effectively support the reliability design of pile foundation.
4th Asia-Pacific Conference on Physical Modelling in Geotechnics (ACPMG2024)
Other