Estimation of Bearing Capacity of Piles in Cohesionless Soil Using Optimised Machine Learning Approaches

Accurate estimation of the bearing capacity of piles requires complex modelling techniques which are not justified by timeframe, budget, or scope of the projects. In this study, six advanced machine learning algorithms including decision tree, k-nearest neighbour, multilayer perceptron artificial neural network, random forest, support vector regressor and extremely gradient boosting are employed to model the bearing capacity of piles in cohesionless soil, and the particle swarm optimisation algorithm is used to optimate the hyper-parameters of machine learning algorithms. A dataset comprising of 59 cases is employed and the R-squared value, root mean square error and variance accounted for are used as performance metrics to compare the performance of optimised machine learning methods. The comparison reveals that the optimised machine learning methods have great potential to estimate bearing capacity of piles and the particle swarm optimisation algorithm is efficient in the hyper-parameter tuning. The results show that R-squared values of six optimised machine learning approaches on the testing set vary from 0.731 to 0.9615. Also, the optimised extremely gradient boosting (R-squared value = 0.9615) shows the best performance compared with other algorithms. Furthermore, the relative importance of influential variable is investigated, which shows that effective stress is the most influential variable for bearing capacity of piles with an importance score of 30.9%. In addition, the results by the optimised machine learning method are compared to the β-method which is a popular empirical method. It is revealed the prominent performance of optimised machine learning approaches.