Artificial intelligence-assisted prediction of critical factors governing pesticide adsorption by agricultural wastes in wastewater treatment

Agricultural waste-based adsorbents offer a sustainable, cost-effective solution for pesticide removal via adsorption. However, predicting removal efficiency under diverse experimental conditions remains challenging, especially when combining quantitative and qualitative variables. This study proposes a robust machine learning framework to predict pesticide removal efficiency using Random Forest (RF), Least-Square Boosting (LSBoost), and Extreme Learning Machine (ELM), each optimized through Bayesian Optimization with six acquisition functions. A curated dataset of 1,983 samples from the literature was used, comprising eight input variables: pesticide type, adsorbent type, initial concentration, contact time, pH, adsorbent to solvent ratio, pretreatment type, and adsorbent form. The target variable was pesticide removal efficiency (%). To determine optimal input combinations, 255 RF-based models (1–8 variables) were developed using cross-validation. RF consistently outperformed other models, achieving the highest predictive accuracy (R = 0.970, NSE = 0.937) and lowest error (NRMSE = 0.099, MAPE = 13.19%). LSBoost produced moderate accuracy, particularly with fewer variables, but showed sensitivity to input structure. ELM underperformed, especially with heterogeneous input features. Feature importance analysis revealed contact time, initial concentration, and adsorbent-to-solvent ratio as the most influential variables, whereas pretreatment type and adsorbent form had minimal impact. Pesticide-specific RF models showed that contact time, adsorbent-to-solvent ratio, and initial concentration were key for atrazine and chlorpyrifos removal, while adsorbent type was also critical for carbofuran. For diazinon, contact time, adsorbent type, and adsorbent-to-solvent ratio were the primary predictors. The findings of this study offer valuable guidance for designing efficient, sustainable pesticide removal systems using agro-waste-based adsorbents. • ML-based prediction of key factors in pesticide removal by agro-waste adsorbents. • Random Forest achieved the highest accuracy and lowest error among tested ML models. • Contact time, concentration, and adsorbent dose were the dominant predictive variables. • Model performance varied with pesticide type and input feature selection. • Study guides eco-friendly design of pesticide removal systems using ML.