The cohesion and internal friction angle of loess are important macro-mechanical parameters for evaluating the safety and stability of engineering construction. Traditional laboratory measurement methods are time-consuming and difficult to conduct on-site. This study aims to compare the effectiveness of five Machine Learning (ML) methods, namely Random Forest (RF), Support Vector Machine (SVM), Back Propagation Neural Network (BPNN), BPNN optimized by Particle Swarm Optimization (PSO-BPNN) and BPNN optimized by Genetic Algorithm (GA-BPNN), in predicting the macro-mechanical properties of loess. To this end, the study collected data from 89 undisturbed loess samples and 229 remolded loess samples to construct training and testing datasets, and used three correlation analysis methods to analyze the influence of physical parameters on mechanical properties. The study found that the water content has the most significant impact on the mechanical properties of loess. In terms of prediction ability, SVM performs the best among the ML methods used, and the determination coefficient for cohesion of undisturbed loess reaches 0.857. Although the training data is limited, the prediction performance of BPNN is significantly improved after being optimized by PSO or GA. The research results show that ML provides an effective way to study the complex mechanical behavior of loess.
