Optimizing Stabilization of Contaminated Mining Sludge: A Machine Learning Approach to Predict Strength and Heavy Metal Leaching

Forecasting the leaching of heavy metals and the strength of contaminated mining sludge (CMS) with high water content is crucial in soil treatment to prevent ecological risks to the environment and human health. Yet, treating high water content CMS effectively is complex due to numerous influencing factors such as binder dosage, water content, curing time, etc. Machine learning has demonstrated significant promise in predicting unconfined compressive strength (UCS) and heavy metal leaching by including various characteristics of CMS stabilized with lime-activated slag. This study assesses the effectiveness of tree-based models such as extreme gradient boosting (XGBoost), decision tree (DT), and light gradient boosting (LightGBM) in predicting the leachability of heavy metal and UCS of treated CMS samples. The dataset consists of 337 samples, of which 80% are designated for training and 20% are designated for testing. Monte Carlo simulations and tenfold cross-validation techniques are applied to improve the model’s performance. The analysis showed that the XGBoost model has superior reliability and accuracy compared to the DT and LightGBM models, with LightGBM showing the lowest predictive performance. The models are ranked as LightGBM < DT < XGBoost based on evaluation metrics such as Adjusted R-squared (Adj. R²), mean absolute error, and root mean squared error. Analysis of feature significance indicates that curing time, GGBS + CaO content, and water content significantly influence UCS. The correlational matrix demonstrates the effect of different input parameters on zinc metal leachability. The leachability of zinc is positively associated with water content but negatively associated with cement content and curing time.