Machine learning prediction of biochar structure stability and adsorption efficiency based on biomass characteristics and aging factors

Abstract

Investigating the biochar aging process can effectively reduce the costs of water and soil remediation while minimizing secondary pollution. Because of the heterogeneity of soil contamination and local climatic conditions, a machine learning framework was developed to circumvent repetitive experimentation and prolonged testing cycles. In this study, a multi-layer nested model was constructed to capture the complex interactions among multiple factors and improve predictive performance. The influences of modeling strategies and database construction on prediction accuracy were systematically evaluated, elucidating the coupling between the physicochemical properties of biochar and aging factors, and subsequently validated against experimental observations. The results revealed three main findings: (1) the random forest model exhibited superior predictive capability for biochar aging, achieving feature correlations of 0.90–0.99 and experimental R² values of 0.80–0.96; (2) biochar structural stability was optimized when carbon, oxygen, and hydrogen contents were maintained at 60–80 %, 10–30 %, and 4–6 %, respectively; and (3) adsorption performance displayed a unimodal trend during coupled freeze–thaw and dry–wet cycles, peaking at 35–40 and 25–35 cycles, respectively, and was further enhanced by adjusting pH, ash content, and elemental ratios. These insights offer valuable guidance for designing and applying biochar in sustainable environmental remediation.