Evaluation of compression index of red mud by machine learning interpretability methods

Abstract

The annual increase in red mud emissions necessitates the expansion of bauxite residue disposal areas (BRDAs), while the escalating land value has led to proposals for development on closed BRDAs. Therefore, understanding the compressive properties of red mud is critical for the safe management and construction of BRDAs. The process of deriving compression index (C_c) through consolidation tests to assess compression characteristics is both time-intensive and vulnerable to the quality of the sampling methods employed. Consequently, it is essential to develop predictive models for compression indices that utilize more easily measurable physical parameters. This study proposes the use of machine learning(ML) models to predict the C_c of red mud. Several machine learning models were studied, including Linear Regression (LR), Ridge Regression (RR), Support Vector Machines (SVR), Random Forest(RF), Extremely Randomized Trees (Extra Trees), K-Nearest Neighbors (KNN), Category Boosting (CatBoost), and LightGBM (Light Gradient Boosting Machine). The grid search algorithm was used to obtain the optimal parameters for each ML model, and k-fold cross-validation was employed to enhance the model’s generalization performance. Ultimately, the KNN model achieved the best performance. The SHAP method was used to describe the specific influence patterns, and to provide a quantitative contribution of each feature to the C_c of red mud. The research indicated that the I_L and w_n exerted the most substantial influence on the C_c, yielding a positive effect.