A Framework for Estimating Matric Suction in Compacted Fine‐Grained Soils Based on a Machine Learning‐Assisted Conceptual Model

Abstract

In this study, a hybrid black‐ and white‐box machine learning (ML) framework is proposed for matric suction estimation in compacted fine‐grained soils, utilizing ML techniques, namely, particle swarm optimized support vector regression (PSO‐SVR) and multi‐gene genetic programming (MGGP). This objective is achieved through developing a novel ML‐based method for designing requisite soil parameters, including a new parameter, the “effective degree of aggregation”. This parameter captures the influence of varying soil structures associated with different initial water content conditions in compacted fine‐grained soils for estimating matric suction. Additionally, sensitivity analyses are performed to better understand the significance of the effective degree of aggregation and other critical soil properties. Explicit equations are derived from the MGGP models, enabling their use for matric suction estimation using spreadsheets and alleviating the reliance on complex programming tools. The proposed models are promising for use in the prediction of hydro‐mechanical behavior and other related compacted fine‐grained soil properties, facilitating their application in conventional geotechnical engineering practice.