Biomethane yield of delignified groundnut shell: A SHAP-based interpretability, clustering, and predictive modeling

Biomethane optimization and upscaling are major challenges in biorefinery due to the recalcitrant properties of lignocellulose biomass and non-linearities of anaerobic digestion. This study integrates the experimental investigation of the influence of Natural Deep Eutectic Solvents (NADES) pretreatment on the biomethane yield of groundnut shells and advanced data-driven techniques encompassing k-means clustering, SHapley Additive exPlanations (SHAP)-based feature interpretations to overcome the black-box nature of conventional machine learning (ML) models and ensemble learning-based predictive modeling. Ethyl glycerol combined with choline chloride at a 1:1 ratio was applied to groundnut shells for 60 min using temperatures of 80 °C and 100 °C, and solid-to-liquid ratios of 1:2 and 1:4 before anaerobic digestion. It was observed that all the pretreatment conditions considered improved the biomethane yield, and the optimum biomethane yield of 365.70 mL CH₄/gVS_added was recorded. Optimum pretreatment conditions were 1:1 of ethyl glycerol and choline chloride for 60 min at a 1:2 solid-to-liquid ratio at 100 °C. SHAP-based feature importance assessment revealed retention times as the dominant driver, with an average SHAP value of about 70. K-means clustering revealed 3 distinct operational clusters reflecting the low, moderate, and high-yield digestion scenarios. XGBoost exhibits a superior predictive performance among others with RMSE, MAE, MAD, MAPE, VAF, and R² values of 0.9641, 0.6614, 0.3576, 9.6905, 99.9, and 0.999, respectively, at training. This research demonstrates that NADES pretreatment, along with SHAP-based interpretable predictive modeling and clustering, can substantially improve biomethane recovery and provide actionable insights for the expansion of lignocellulosic bioenergy systems.