Machine learning framework to predict product distribution of lignocellulosic biomass pyrolysis

Area of Science:

• Chemical Engineering
• Computational Chemistry
• Materials Science

Background:

• Biomass pyrolysis is complex, making first-principles modeling challenging.
• Machine learning offers an alternative but faces data scarcity and interpretability issues.
• Interpretable models are crucial for understanding and optimizing biomass conversion.

Purpose of the Study:

• Develop an interpretable machine learning framework for biomass pyrolysis.
• Utilize data from fixed-bed lignocellulosic biomass pyrolysis experiments.
• Enhance understanding and decision-making in biomass pyrolysis processes.

Main Methods:

• Proposed a mass change basis for constructing machine learning models.
• Implemented Artificial Neural Network (ANN) and Symbolic Regression (SR) models.
• Assessed feature importance using SHAP and PLS, with PLS identifying optimal features for SR.

Main Results:

• Both ANN and SR models achieved R² > 0.85 for all phase products in the testing set.
• SR models showed superior generalization (R² > 0.9) in extrapolation tests for char and gas phases.
• Uncertainty assessment improved SR model robustness and prediction stability.

Conclusions:

• The developed framework provides a valuable, interpretable tool for biomass pyrolysis modeling.
• The approach aids in understanding complex chemical processes and supports decision-making.
• Symbolic regression shows promise for robust biomass pyrolysis prediction, with limitations for high oil yields.