Efficient machine learning method for evaluating compressive strength of cement stabilized soft soil

Abstract

To meet the design strength of cement stabilized soft soil (cement soil) in different application environments, numerous field and indoor geotechnical tests are often conducted. These tests not only result in the unreasonable use of materials, cost and time, but also cause serious environmental pollution problems. This study aims to propose an efficient machine learning model to evaluate the compressive strength of cement soil. Firstly, a database containing 566 samples was developed by literature collection. Secondly, eight machine learning models were established to train and test the database, as well as evaluated for the generalization ability using six performance indicators. Finally, the optimal model was selected to conduct the correlation importance analysis of feature variables using shaply additve explanations and partial dependence plots, and compared with the typical empirical model. The results indicated that the compressive strength of cement soil was better predicted by extreme gradient boosting model (the determinate coefficient of model in the test set was 0.93). The cement content, water content, curing age and fine grain were the main feature variables influencing the compressive strength. A reliable database and machine learning model are proposed, which have practical reference for the design and construction of cement soil in soft foundation projects.