Unified interpretable machine learning framework for predicting pellet quality from raw and thermochemically...

Biomass pellets offer a renewable pathway for decarbonising industrial energy and metallurgical processes, yet inconsistent quality limits widespread adoption. This study presents a unified, interpretable machine learning framework that predicts pellet density and mechanical strength across raw biomass, hydrochar, and torrefied biomass feedstocks. A literature-derived dataset spanning feedstock properties, pretreatment conditions, and densification parameters was compiled to capture heterogeneity in pelletisation systems. Eight algorithms were benchmarked using Bayesian optimisation and 5-fold cross-validation, achieving R² > 0.85 with root-mean-square errors within experimental uncertainty. Interpretability analyses revealed critical nonlinear interactions among binder content, feedstock composition, and thermo-mechanical conditions that influence densification performance. The unified framework matched experimental-grade pellet specifications (∼1.2 g cm^-3; 6-7 MPa) while reducing computational cost by approximately 25% compared to pathway-specific models. The framework provides generalisable thermo-mechanical design principles and is deployed through a graphical interface, enabling users to predict pellet quality based on compositional and process inputs.