Machine learning prediction of biochar yield based on biomass characteristics

Area of Science:

• Biomass conversion
• Thermochemical processes
• Materials science

Background:

• Slow pyrolysis is a key thermochemical process for converting organic waste into valuable biochar.
• Optimizing biochar yield requires understanding complex interactions between process parameters and feedstock characteristics.

Purpose of the Study:

• To develop predictive machine learning models for biochar yield.
• To identify key factors influencing biochar production during slow pyrolysis.
• To assess the performance of different machine learning algorithms in this predictive task.

Main Methods:

• Employed six machine learning models, including gradient boosting decision trees and neural networks.
• Utilized Pearson feature selection to identify relevant variables.
• Applied partial dependence analysis to explore feature-variable relationships.

Main Results:

• Gradient boosting and Levenberg-Marquardt neural networks achieved high predictive accuracy (R² > 0.9 training, R² > 0.8 testing).
• Partial dependence plots indicated pyrolysis conditions, especially temperature, have a greater impact on biochar yield than biomass composition.
• Highest treatment temperature emerged as a critical, consistently adjustable factor for regulating biochar yield.

Conclusions:

• Machine learning offers a powerful tool for predicting experimental outcomes in pyrolysis.
• This approach can significantly reduce time and economic costs in biochar production research and development.
• Findings provide a scientific basis for optimizing slow pyrolysis processes for enhanced biochar yield.