Mixture-of-Experts Machine Learning Framework for Predictive Design of Biomass-Derived Hydrochar to Decarbonize...

Area of Science:

• Renewable Energy Engineering
• Materials Science
• Artificial Intelligence

Background:

• Hard-to-abate industries like steel and cement require high-temperature heat, necessitating predictable renewable solid fuels.
• Biomass residues can be converted to hydrochar via hydrothermal carbonization (HTC), but feedstock variability hinders property prediction.
• Current methods struggle to reliably predict hydrochar properties and assess emissions reduction potential due to process variability.

Purpose of the Study:

• To develop a machine learning framework for accurately predicting hydrochar properties (Higher Heating Value - HHV, Energy Yield - EY).
• To overcome limitations posed by feedstock heterogeneity and process variability in HTC.
• To enable data-driven design of hydrochar for industrial decarbonization.

Main Methods:

• Implemented a Mixture of Experts (MoEs) machine learning strategy.
• Integrated clustering algorithms with tailored regression models and a gating network.
• Utilized multiobjective optimization to identify optimal HTC conditions for maximizing HHV and EY.
• Performed experimental validation to confirm model robustness.

Main Results:

• Achieved superior accuracy in predicting critical hydrochar properties (HHV and EY).
• Identified optimal HTC conditions for wood chips, corn straw, and sludge, maximizing both HHV and EY.
• Demonstrated experimental validation confirming the model's robustness and predictive power.

Conclusions:

• The MoEs framework provides a data-driven paradigm for scalable hydrochar design.
• Optimally produced hydrochar can achieve net energy gains and significant CO2 emission reductions.
• Potential to reduce China's annual CO2 emissions by 3.3% (396.6 million tons) through deployment across agricultural residues and municipal sludge.