<p>Torrefaction is an effective thermochemical pretreatment for upgrading agricultural and forestry biomass into solid biofuels; however, selecting operating conditions remains challenging due to trade-offs between mass yield and energy densification. In this study, a comprehensive machine-learning framework was developed to predict biomass mass yield and higher heating value using an experimental dataset collected from diverse torrefaction studies. Multiple regression algorithms—including linear, ensemble, boosting, and kernel-based models—were systematically compared with and without hyperparameter tuning. Domain-informed feature engineering and rigorous data preprocessing were applied to enhance predictive reliability and interpretability. Among the evaluated models, tree-based ensemble and boosting algorithms, particularly CatBoost, showed robust performance and stable feature attribution for both MY and HHV. Rather than treating prediction accuracy as an endpoint, the models were integrated into a GUI-based simulator that enables simultaneous evaluation of MY and HHV under identical operating conditions, visualizes trade-offs, and identifies feasible operating windows without exhaustive trial-and-error. This approach demonstrates how ML can function as a practical engineering tool for torrefaction screening and condition selection, complementing traditional empirical methods.</p>