Energy and carbon routing in hydrothermal conversion of lignocellulosic biomass: An uncertainty-aware machine-learning framework

Reliable assessment of hydrothermal lignocellulosic biomass conversion requires quantitative prediction of how feedstock mass, chemical energy, and carbon partition among bio-oil, solid residues, aqueous products, and gases under heterogeneous feedstock, solvent, catalyst, and severity conditions. We constructed a harmonized database of 3,693 HTL/HTC experiments from 187 scientific papers and 433 different biomass samples, with temperatures up to ∼ 600 ◦C and maximum residence times of up to 1,200 min. From this literature-scale dataset, we developed an uncertainty-aware machine-learning framework for predicting mass partitioning, energy and carbon allocation, and higher heating value (HHV) of condensed phases. Among the tree-ensemble regression methods examined, Random Forest (RF) provided the most stable cross-target compromise, with test R2 values for product yields around 0.8, energy/carbon allocation metrics between 0.8 and 0.9, and bio-oil HHV around 0.7 with root-mean-squared error ∼ 2–3 MJkg−1. The model recovers established severity, solvent, and composition trends while identifying regions where predictions are well supported or weakly constrained. SHAP attributions are interpreted as model-conditional associations: liquid-phase outcomes depend strongly on solvent and operating conditions, whereas solid-phase HHVs are more associated with feedstock carbon and ash content. Ensemble and conformal uncertainty methods distinguish relative epistemic uncertainty from calibrated scenario bounds. The framework is intended for uncertainty-aware process screening, scenario bounding, and experimental prioritization rather than unqualified point optimization in poorly represented regimes.