Application of Ensemble-Based Methods for Prediction of Undrained Shear Strength of Soft Sensitive Clays

Owing to recent advancements in artificial intelligence, the technology has gained popularity across all possible domains in geotechnical engineering over the past few decades. The present study demonstrates the application of machine learning techniques to predict the undrained shear strength of soft, sensitive clay. Field and laboratory measurements from 24 test locations in Finland are used to create a multivariate database consisting of 384 data points. The main objective of the study is to predict the undrained shear strength using ensemble-based machine learning models, viz., random forest, gradient boosting, extreme gradient boosting, AdaBoost, and light gradient boosting. The undrained shear strength of the clay soil sample is considered as an output parameter for given input features that include pre-consolidation stress, vertical effective stress, liquid limit, plastic limit, and water content. The results from the ensemble-based machine learning models are compared with those from the base models like decision trees. Results show that the ensemble-based extreme gradient boost and light gradient boost models outperform the base models in terms of accuracy and error with the extreme gradient boost model showing the highest coefficient of determination and root mean squared error (indicators of a good prediction model). Furthermore, sensitivity analysis is carried out for the best-performing extreme gradient boosting model, and results show that undrained shear strength is highly influenced by pre-consolidation stress, the existing vertical effective stress, and the plasticity index of the soil.