A Comparative Analysis of Machine Learning Models for Predicting Loess Collapse Potential

Collapsible soils, particularly loessial soils, present significant geotechnical engineering hazards that should be carefully investigated before any construction can commence. However, it is generally difficult to estimate the collapse potential of soils based on the relative contributions of each of the numerous influencing factors. Therefore, the main objective of this study is to find a reliable method for predicting the collapse potential of loessial soils by using machine learning-based tools. In this regard, details of 766 performed oedometer test were gathered from the published literature containing six variables for each data point including dry unit weight of soil, plasticity index, void ratio, degree of saturation, inundation stress at which the oedometer test was conducted, and the collapse potential. Then, prediction for the degree of collapsibility of loess was performed by employing three well-known supervised machine learning tools, namely Multi-Layer Perceptron Neural Network (MLPNN), Radial Basis Function Network (RBFN), and Naive Bayesian Classifier (NBC), and outcomes were analyzed based on a comparative view. Simulation results indicate the superiority of MLPNN in estimating the degree of collapsibility of loess against other models in terms of performance error metrics and precision criterion.