Machine learning-driven modeling of biomass pyrolysis product distribution through thermal parameter sensitivity

This study investigated the application of machine learning to simultaneously predict and optimize biomass pyrolysis product yields (bio-oil, bio-char, gas) using practical input parameters. A dataset of 273 experiments from 38 literature sources, encompassing diverse agricultural feedstocks and varying pyrolysis conditions (temperature and heating rate), was compiled. Five machine learning models, including Multiple Linear Regression (MLR), k-Nearest Neighbors Regression (kNN), Support Vector Regression (SVR), Random Forest Regression (RFR), and XGBoost Regression (XGB), were evaluated for predicting yields. XGB significantly outperformed other models, achieving a training R2 over 0.900. Feature importance analysis identified key parameters influencing product yields: cellulose and hemicellulose for bio-oil, temperature and lignin for bio-char, and lignin and hemicellulose for gas. Subsequent sensitivity analysis, varying temperature and heating rate across cellulose-rich, hemicellulose-rich, and lignin-rich biomass, revealed distinct product distribution trends. Comparison with experimental data validated the model's accuracy with an overall R2 of 0.900. This work demonstrates the potential of XGB, using practical input parameters, to develop customized pyrolysis guidelines for optimizing product yields from specific feedstocks, contributing to efficient biomass utilization.