Interpretable data-driven insights into catalytic hydrothermal liquefaction: trade-offs between denitrogenation and energy recovery

This study develops an interpretable data-driven framework for catalytic hydrothermal liquefaction (HTL) of biomass, with emphasis on the trade-off between denitrogenation and energy recovery. Stratified kernel density estimation, interaction analysis, machine learning benchmarking, and game-theory-based feature interpretation were integrated to examine how feedstock composition, reaction conditions, and alkali catalyst descriptors are jointly associated with variations in biocrude elemental quality, nitrogen retention, and energy recovery. The results show that HTL performance is highly structured but strongly nonlinear. Feedstock elemental composition and operating conditions are the principal model predictors delimiting the main performance envelope, while catalyst descriptors are associated with variations in the balance between nitrogen removal and energy preservation. Hydroxide catalysts with stronger basicity are accompanied by greater denitrogenation under severe conditions, but often reduced energy recovery. In contrast, carbonate catalysts showed more balanced behavior, with K 2 CO 3 providing within the compiled literature dataset the most favorable compromise between nitrogen control and energy preservation. Across prediction tasks, tree-based ensemble models achieved the best balance between predictive accuracy and generalization. Linear models underfitted the data, support vector regression tended to overfit, and multilayer perceptron performance was less stable. CatBoost emerged as the most consistent unified model for predicting elemental ratios, nitrogen retention, and energy recovery. Model-attributed predictive contributions identified catalyst loading, molecular mass, and pH as the dominant catalyst descriptors. This work explicitly resolves model-predicted patterns in the catalyst-associated trade-off between nitrogen retention and energy recovery across multiple biocrude-quality indicators, providing a practical framework for catalyst screening and HTL process optimization.