Application of multiple machine learning models integrated with SHAP analysis in predicting the compressive strength of molybdenum tailings concrete

Traditional trial mixing methods for determining the compressive strength of molybdenum tailings concrete are characterized by long cycles, high costs, and susceptibility to external interference. Therefore, this study proposes the application of machine learning methods to predict the compressive strength of molybdenum tailings concrete, aiming to reduce ineffective trial-and-error experiments while providing an efficient and reliable new approach for strength prediction. Through systematic collection of relevant literature, this study constructed a dataset comprising 104 sets of molybdenum tailings concrete data, covering six input features (stone, cement, water, water-reducing agent, and molybdenum tailings per cubic meter of concrete) and one output feature (compressive strength). Ten machine learning models—Linear Regression, Ridge Regression, K-Nearest Neighbors, Support Vector Machine, Decision Tree, Multilayer Perceptron, Random Forest, Adaboost, Gradient Boosting, and XGBoost—were employed to predict the compressive strength of molybdenum tailings concrete. The optimal model was selected through statistical quantitative evaluation and dispersion analysis, and the SHAP method was used to interpret the influence of each input variable on the compressive strength of molybdenum tailings concrete. The results show that the MLP model exhibited the best predictive performance, with evaluation metrics on the test set as follows: RMSE = 3.027, MAE = 2.279, MAPE = 4.442, R ² = 0.918, achieving the highest comprehensive score. Moreover, the four models—MLP, Adaboost, Gradient Boosting, and XGBoost—demonstrated the smallest dispersion and all fell within the acceptable error range. Additionally, SHAP analysis further indicated that cement content is the core parameter influencing compressive strength, and when the molybdenum tailings replacement rate is in the range of 20%–25%, its negative impact on compressive strength is minimized. This conclusion aligns with existing engineering practices and experimental findings.