Water pollution from pesticides is a significant environmental problem, particularly in developing nations like India and China, hindering the achievement of UN Sustainable Development Goals. Rapid urbanization and industrialization exacerbate the issue, threatening biodiversity and human health. Pesticides contaminate surface and groundwater, entering the food chain and causing various health problems. In-situ remediation using biochar adsorption is a promising sustainable and cost-effective alternative. Biochar, produced from various sources through different methods, shows potential for removing various pesticides from aqueous environments. However, its efficiency is influenced by numerous factors: feedstock, production conditions, water matrix (pH, concentration, temperature), experimental conditions (dose, time), and pesticide properties (solubility, molecular weight). Numerous adsorption mechanisms are at play. Despite extensive research, optimal conditions for biochar-mediated pesticide removal remain variable. Review studies and bibliometrics are time-consuming. This research proposes a computational approach using machine learning to optimize biochar use, reducing time, cost, and resource needs. Machine learning models, such as support vector machines (SVM), convolutional neural networks (CNN), random forests (RF), and artificial neural networks (ANN), have proven useful in pesticide analysis, tracking dissipation, assessing soil microbial impacts, and evaluating genotoxic effects. However, the application of ensemble ML algorithms to predict the efficacy of biochar adsorption for pesticide remediation remains unexplored. This study uses CatBoost, LightGBM, and RF to predict the adsorption efficiency considering biochar properties, aqueous matrix configuration, and experimental conditions.

The introduction extensively reviews existing literature on pesticide pollution, the health risks associated with pesticide exposure, existing remediation techniques, and the potential of biochar for pesticide removal. It highlights the challenges in determining optimal biochar application conditions due to the interplay of various factors and the limitations of traditional review methods. The literature supports the need for a data-driven approach using machine learning to predict and optimize the effectiveness of biochar for pesticide remediation.

This study compiled data from 96 research articles on biochar-mediated pesticide adsorption from various online databases (Google Scholar, Scopus, Web of Science). Data on pesticide adsorption capacity, biochar properties (surface area, pore volume, pH), water matrix parameters (pH, temperature, initial pesticide concentration), and experimental conditions (biochar dose, contact time) were extracted. Missing data were imputed using the median value after removing attributes with more than 50% missing values. Three ensemble machine learning models—CatBoost, LightGBM, and Random Forest (RF)—were trained and evaluated using a stratified random sampling approach (70% training, 30% testing). Hyperparameter tuning was performed using grid search to minimize root-mean-squared error (RMSE). Model performance was assessed using the coefficient of determination (R²) and RMSE. Feature importance was analyzed using CatBoost's built-in feature importance function and SHAP (SHapley Additive exPlanations) values to understand the relative contributions of each input attribute to pesticide adsorption. Partial dependence plots (PDP) were used to visualize the relationships between individual and combined attributes and adsorption capacity.

The CatBoost model showed the best performance with R² values of 0.968 and 0.956 for the training and testing sets, respectively, outperforming LightGBM and RF. Feature importance analysis revealed that surface area (SA) was the most significant factor influencing pesticide adsorption, followed by pesticide concentration (Co) and pore volume (Vt). Partial dependence plots showed a positive correlation between SA and Vt and adsorption capacity within specific ranges. Increasing biochar dose initially increased adsorption but then decreased it due to pore blockage or reduced surface area per unit mass. Optimal biochar dose was found to be less than 1 g/L with a treatment time less than 500 minutes. Pesticide concentration had a linear positive relationship with adsorption capacity, reflecting increased mass transfer at higher concentrations. The model also indicated that biochar pH had the least influence among the input parameters. The SHAP values provide a visualization of the individual feature impact on model predictions and further highlight the relative importance of the various parameters.

The superior performance of CatBoost in predicting pesticide adsorption highlights the potential of machine learning for optimizing biochar application. The identified key factors—surface area, pesticide concentration, and biochar dose—provide valuable insights for designing and producing biochar with enhanced adsorption capacity. Optimizing biochar production to maximize surface area and pore volume, combined with the understanding of the relationship between biochar dose and treatment time, can lead to more efficient pesticide removal. The results suggest a pathway towards sustainable agriculture by optimizing resource utilization, enhancing soil health through biochar amendment, and mitigating environmental contamination. While the training data primarily came from laboratory-scale experiments, integrating real-world data will further enhance the model's generalizability and predictive power. Future research could explore the post-adsorption management of saturated biochar.

This study demonstrates the successful application of ensemble machine learning, specifically CatBoost, to predict the efficiency of biochar in removing pesticides from water. The model highlights the importance of biochar textural properties, pesticide concentration, and biochar dosage in the adsorption process. The findings provide a data-driven approach for optimizing biochar design and use, promoting sustainable agricultural practices and environmental remediation. Future work should focus on incorporating real-world data to improve model generalizability and exploring ML-assisted management of saturated biochar and desorbed pesticides.

The study's limitations include the reliance on data from laboratory-scale experiments, which may not perfectly reflect real-world conditions. The model's generalizability could be improved by incorporating data from diverse agricultural settings and environmental conditions. The analysis focused primarily on adsorption efficiency, and further research could investigate other aspects of biochar's role in pesticide remediation, such as desorption and long-term environmental impact. The data compilation relied on existing literature, and inconsistencies in experimental methods and reporting across different studies could introduce bias.