Using machine learning approaches to construct correlations for cohesive soils usins using in-situ and laboratory data

Abstract

In a world where the sizes of construction sites are ever increasing and project deadlines ever reducing, the geotechnical engineer no longer has the time to properly conduct the necessary tests on the soil so as to come up with optimal soil properties that would as accurately as possible re ect the ones on site. Therefore, correlations equations together with in-situ tests and laboratory tests have formed the basis of geotechnical engineering design. The literature is lled with correlation equations developed by previous and present researchers. Some of these equations may or may not have any statistical background hence making them less reliable when used to estimate critical soil parameters. The goal of any correlation equation developed is to estimate as accurately as possible a response given a certain input. In this thesis, we aim at developing regression models using machine learning algorithms such as linear regression, Random Forest and Gradient Boosting so as to predict the undrained shear strength, cu, the elastic modulus, Em and the limit pressure, pL. In order to further improve our prediction capabilities we can stack the aforementioned models using their weighted averages derived from their RMSE indices obtained from the test data. Finally, the best performing models are compared to the correlations equations found in the literature.