Enhancing lead adsorption capacity prediction in biochar: a comparative study of machine learning models and parameter optimization.

Abstract

Machine learning models for predicting lead adsorption in biochar, based on preparation features, are currently lacking in the environmental field. Existing conventional models suffer from accuracy limitations. This study addresses these challenges by developing back-propagation neural network (BPNN) and random forest (RF) models using selected features: preparation temperature (T), specific surface area (BET), relative carbon content (C), molar ratios of hydrogen to carbon (H/C), oxygen to carbon (O/C), nitrogen to carbon (N/C), and cation exchange capacity (CEC). The RF model outperforms BPNN, improving R2 by 10%. Additional features and particle swarm optimization enhance the RF model's accuracy, resulting in an 8.3% improvement in R2, a decrease in RMSE by up to 56.1%, and a 55.7% reduction in MAE. The importance ranking of features places CEC > C > BET > O/C > H/C > N/C > T, highlighting the significance of CEC in lead adsorption. Strengthening the complexation effect may improve lead removal in biochar. This study contributes valuable insights for predicting and optimizing lead adsorption in biochar, addressing the accuracy gap in existing models. It lays the foundation for future investigations and the development of effective biochar-based solutions for sustainable lead removal in water remediation.