Interpretable machine learning to model biomass and waste gasification.

Area of Science:

• Thermochemical processes
• Chemical engineering
• Machine learning applications

Background:

• Machine learning (ML) offers potential for modeling complex thermochemical processes like gasification.
• The 'black box' nature of ML models can hinder trust and understanding.
• Accurate modeling is crucial for optimizing biomass and waste gasification.

Purpose of the Study:

• To evaluate seven ML methods for modeling biomass and waste gasification.
• To identify key factors influencing gasification outcomes.
• To enhance the interpretability of ML models in this domain.

Main Methods:

• Application of seven distinct ML algorithms to gasification data.
• Utilizing global and local interpretability techniques (e.g., SHAP, LIME).
• Performance evaluation using coefficient of determination (R²) across ten gasification outputs.

Main Results:

• Gradient boosting regression demonstrated superior performance with an average R² of 0.90.
• Feedstock particle size and gasifying agent type were identified as primary influential factors.
• Interpretability methods successfully illuminated model behavior and key drivers.

Conclusions:

• ML models, particularly gradient boosting, can accurately predict gasification outputs.
• Enhanced model interpretability improves understanding of the gasification process.
• Informed decisions regarding gasification process design are facilitated for stakeholders.