Prediction of methane production in biochar-assisted anaerobic digestion using automated machine learning with biochar properties, environmental factors, and methanogens

초록

Biochar has been widely applied in anaerobic digestion (AD) to enhance process efficiency and stability. However, most existing predictive studies of methane yield in AD have overlooked microbial community data, limiting their biological interpretability. In this study, automated machine learning (AutoML) was employed to predict methane yield in biochar-assisted AD systems using a multi-study dataset integrating biochar properties, operating conditions, and the relative abundance of key methanogens. Seven feature set combinations were systematically evaluated to identify the most effective predictors. The integrated feature set (biochar + environment + methanogens) and the environment-methanogen combination achieved the highest predictive performance, with coefficients of determination (R2) of 0.79 and 0.68 and root mean square errors (RMSE) of 24.87 and 28.85, respectively. The performance of the biochar-only model (R2 = 0.16, RMSE = 101.24) improved markedly to R2 = 0.50 and RMSE = 53.4 when methanogen features were added (biochar + methanogens). Accuracy increased to R2 = 0.79 and RMSE = 24.87 with the integration of all feature categories, highlighting the substantial predictive value of microbial information. Among eleven AutoML-generated models, CatBoost outperformed others due to its higher capacity to handle nonlinear and high-dimensional data. Feature importance analysis identified hydraulic retention time, temperature, Methanomassiliicoccus, and Methanosaeta as key factors influencing methane production. This study demonstrates the importance of methanogenic data and the effectiveness of AutoML in modeling complex biochar-assisted AD systems. The findings further provide actionable insights for optimizing methane yield in both laboratory and full-scale applications.

키워드