Abstract:

The prediction methods for the collapsibility coefficient of loess are predominantly based on regression analysis, but they are significantly influenced by the regional characteristics of loess, leading to insufficient prediction accuracy. Furthermore, existing machine learning prediction methods have limitations in interpretability. Focusing on the loess in the Longxi region, five pivotal influencing factors, namely, density, saturation, self-weight stress, liquid limit, and internal friction angle, were identified through correlation and cluster analyses. Subsequently, prediction models for the loess collapsibility coefficient were constructed using extreme gradient boosting (XGBoost), random forest (RF), support vector machine (SVM), and regression analysis. Additionally, an interpretability analysis of the prediction outcomes was performed using the Shapley additive explanations (SHAP) technique. The results show that the correlation coefficients between the collapsibility coefficient of Longxi loess and various indices such as density, dry density, void ratio, saturation, and internal friction angle range from 0.63 to 0.90, indicating extremely strong or strong correlations. Both the XGBoost and RF models demonstrate a prediction accuracy exceeding 90% for the collapsibility coefficient and collapsibility grade, thereby outperforming the SVM model and conventional regression methods. The SHAP analysis reveals that loess with a density less than 1.44 g/cm³ has a higher probability of collapsibility, while loess with a saturation greater than 29.0% has a relatively lower collapsibility grade. Moreover, loess under high saturation and high self-weight stress conditions exhibits weaker collapsibility tendency and lower collapsibility grade. The findings provide a theoretical reference for loess engineering construction in the Longxi region.

Key words: Longxi loess, collapsibility coefficient, parameter correlation analysis, machine learning, SHAP interpretability analysis