Kinetic parameter estimation for biomass pyrolysis: experimental investigation and automated machine learning prediction

This study establishes an integrated framework that combines multi-method experimental kinetics with interpretable automated machine learning for predicting and elucidating the activation energy of biomass pyrolysis. Thermogravimetric analysis and kinetic modelling of three representative feedstocks (poplar, pine, corn straw) were performed using Kissinger–Akahira–Sunose, Flynn–Wall–Ozawa, Starink, Friedman, and Coats–Redfern methods. Gaussian multi-peak deconvolution clarified the stepwise decomposition of pseudo-components, revealing the thermal stability order: lignin > cellulose > hemicellulose. A robust, interpretable machine learning model was developed using the Fast Library for Automated Machine Learning based on a curated dataset of 1749 samples from 59 studies. The model achieved high prediction accuracy (test set: coefficient of determination = 0.74–0.89, Root Mean Square Error = 19.01–27.75 kJ/mol) and identified conversion degree, nitrogen, hydrogen, and lignin content as the dominant features governing activation energy. SHapley Additive exPlanations analysis, three-dimensional partial dependence plots, and interaction analysis further quantified nonlinear feature interactions, revealing that high hydrogen and lignin contents synergistically elevate activation energy, whereas high oxygen content mitigates lignin's stabilizing effect. This work bridges component-specific kinetic analysis with explainable machine learning, providing both a predictive tool and mechanistic insights for the targeted optimization of biomass pyrolysis.