Simulation, prediction and optimization of typical heavy metals immobilization in swine manure composting by using machine learning models and genetic algorithm

Abstract

Machine learning (ML) is a novel method of data analysis with potential to overcome limitations of traditional composting experiments. In this study, four ML models (multi-layer perceptron regression, support vector regression, decision tree regression, and gradient boosting regression) were integrated with genetic algorithm to predict and optimize heavy metal immobilization during composting. Gradient boosting regression performed best among the four models for predicting both heavy metal bioavailability variations and immobilization. Gradient boosting regression-based feature importance analysis revealed that the heavy metal initial bioavailability factor, total phosphorus, and composting duration were the determinant factors for heavy metal bioavailability variations (together contributing >75%). After genetic algorithm optimization, the maximum immobilization rates of Cu, Zn, Cd, As, and Cr were 79.53, 31.30, 14.91, 46.25, and 66.27%, respectively, superior to over 90% of the measured data. These findings demonstrate the potential application of ML to risk-control for heavy metals in livestock manure composting.

Keywords: Composting; Genetic algorithm; Heavy metal; Machine learning; Risk reduction.