Machine learning-driven predictive frameworks for optimizing chemical strategies in Microcystis aeruginosa mitigation

Abstract

Traditional chemical approaches to control Microcystis aeruginosa often struggle to define optimal treatment conditions resulting in inconsistent outcomes. Previous research largely depended on single-source data, significantly compromising generalizability of these findings. These studies overlooked comparisons among various machine learning models. To address these limitations, we integrated experimental data from multiple sources for the first time and applied various machine learning models to predict removal efficiency. This research enhances chemical treatment efficiency through data driven optimization of critical factors such as time and dosage required for effective Microcystis aeruginosa removal, removing uncertainties in traditional experimental studies. Another key point of novelty is to extract key features influencing output, which have not been quantitatively explored in earlier studies. The study observed variability in removal time, dosage and efficiency rates, with an average removal efficiency of 50.03 %. Random Forest Regressor and Bagging Regressor were recommended as optimum models, demonstrating their effectiveness in accurately predicting removal efficiency based on the given dataset. Our findings indicate that removal efficiency was most sensitive during initial 0–500 min of treatment and at dosages below 250 mg/L. The optimal dosage range of 0–250 mg/L was identified, with significant drops in removal efficiency beyond this range, indicating potential risks of reduced effectiveness and adverse effects at higher concentrations. The study also underscores the potential of photocatalysts, heterogeneous catalysts, and the chemical Bi₂O₃ in optimizing removal efficiency. This predictive framework provides decision-makers with essential tools for effectively predicting the efficiency of chemical mitigation strategies against harmful algal blooms.