Decarbonizing hard-to-abate industrial sectors that require high-temperature process heat, notably steel and cement, demands renewable solid fuels with rigorously predictable properties. Hydrothermal carbonization (HTC) of biomass residues offers a promising route to such fuels, yet feedstock heterogeneity and process variability impede the reliable prediction of hydrochar properties and emissions reduction potential. Here, we introduce a machine learning framework leveraging a Mixture of Experts (MoEs) strategy to overcome these limitations. Our approach integrates clustering algorithms with tailored regression models and a gating network for autonomous model assignment, achieving superior accuracy in predicting critical hydrochar properties of higher heating value (HHV) and energy yield (EY). Through multiobjective optimization, we identify HTC conditions that simultaneously maximize HHV and EY for wood chips, corn straw, and sludge, with experimental validation confirming model robustness. We further demonstrate that optimally produced hydrochar can deliver net energy gains and reduce CO2 emissions by 396.6 million tons annually if deployed across China's agricultural residues and municipal sludge, equivalent to 3.3% of annual national emissions. This MoEs framework establishes a data-driven paradigm for scalable hydrochar design, enabling the targeted decarbonization of emission-intensive industries.
Guo et al. (Tue,) studied this question.