Machine learning approaches for identifying the influencing factors of input materials in cement and predicting heavy metal concentrations

Abstract

Waste co-processing technologies for producing cement can lead to elevated levels of heavy metals (HMs) and other harmful pollutants, which can negatively impact humans and the environment. We evaluated three pollution indices of HM concentrations in cement, namely pollution load index (PLI), potential ecological risk index (PERI), mean probable effect concentration (m-PEC). According to results, 39–74 % of cements exhibited PLI > 1. According to the m-PEC index, the ecological risk level was in all cases in the range of 52–100 %, i.e., considerable risk. All cements exhibited low PERI values of <150, indicating low ecological risk. A machine learning approach was adopted to identify the influencing factors and predict the heavy metal concentrations in cement, with the Adaptive Synthetic Sampling oversampling technique enhancing the predictive accuracy. Classification models using the artificial neural network (ANN) and random forest (RF) techniques for six heavy metals showed that RF exhibited higher predictive accuracy than that of ANN for all HMs except As. In both the ANN and RF models, the highest predictive accuracy was achieved for Pb, i.e., 0.785 and 0.840, respectively. Excluding Pb, the range of predictive accuracies for heavy metals of the ANN model was 0.583–0.683, whereas that of the RF model was 0.635–0.771. Feature importance calculations showed that As and Cd concentration in cement was mainly influenced by sludge input as a replacement raw material, whereas clay played a significant role in determining the Hg and Pb concentrations in cement. The other heavy metals were influenced by a variety of sources with similar contributions. © 2025