Machine learning-based pedo transfer function for estimating the soil compression index

Abstract

Soil compression index (Cc) plays a vital role in describing the settlement behaviors of geotechnical infrastructures. The conventional Oedometer test broadly used to determine Cc is time-consuming and expensive, which challenges incorporating the high spatial variability of Cc. Alternatively, this study utilized the pedo transfer function (PTF) concept to develop a predictive model on the extreme gradient boosting (XGB) framework for estimating Cc with high accuracy and low effort. The presented XGB-PTF implemented on the database is acquired from 40 boreholes in Ho Chi Minh city and its vicinity to learn and recognize the correlation patterns of Cc and the easily-obtainable soil parameters (i.e., grain size distribution, unit density, moisture content, Atterberg limits). Rigorous evaluation with standard regression metrics demonstrated the efficiency and excellent performance of the XGB-PTF (e.g., low root-mean-squared error of 0.089 and a high coefficient of determination of 0.903). Furthermore, the presented framework showed its superiority over the current empirical equations in estimating Cc by higher prediction accuracy and applicability to the broader range of soil types. Given efficiency, flexibility, and dynamics, the presented model is expected to be a versatile approach to quantizing and advancing the knowledge of soil characteristics over a regional area.