Prediction of Bi-Linear Strength Envelope of Brazilian Soils Using Machine Learning Techniques

The knowledge of soil strength envelope is essential for several works in civil engineering, but laboratory tests to obtain them can be costly. However, based on recently produced studies, it is believed that the development of computational models to estimate them is a tool capable of meeting this demand. This study aims, therefore, to develop a machine learning model capable of estimating bi-linear strength envelopes of soils. In order to achieve it, a database was built with results from characterization, direct shear and Brazilian tests from Brazilian soils. Five regression algorithms were applied to the database, undergoing training, cross-validation, hyperparameter optimization and testing phases. Most models presented satisfactory performance metrics, emphasizing Linear Regression algorithm as the most adequate for direct shear tests and Extremely Randomized Trees algorithm for Brazilian tests.