Use of Multivariate Adaptive Regression Splines (MARS) in the Performance Prediction of Anti-floating Anchors

Fully grouted ground anchors have been increasingly used as a part of foundation system to resist buoyant force in geotechnical practice. However, designs of fully grouted anchors are commonly based on the calculation of the ultimate pullout capacity along with safety factors, which results in unnecessary economic loss. This is partly due to the fact that it is impractical to predict the anchor performance without strong assumptions of how steel tendons, soils, rock, and grout can collectively resist pullout force or without detailed information of the ground parameters. As one of the promising fields within the framework of artificial intelligence, Machine Learning (ML) has been increasingly used to address geotechnical problems by giving computers the ability to learn without being explicitly programmed. Multivariate Adaptive Regression Splines (MARS) is an ML nonparametric algorithm that is based on a data-driven process. This paper presents the development of a MARS performance prediction model using data from 530 anti-floating anchor pullout tests in 8 different projects in weathered soils and rocks located in Shenzhen, China. In this study, MARS demonstrates the capabilities to capture the complex non-linear relationships in the anti-floating anchor pullout problem. In addition, it is shown that the displacement-based design procedure of the anti-floating anchor based on the MARS model is feasible if appropriate safety factors are adopted.