Machine-learning-based cross-prediction of ultimate and proximate properties for cleaner biomass fuel assessment

Reliable assessment of biomass fuel quality is essential for cleaner solid-fuel utilisation and thermochemical upgrading, but repeated ultimate analysis, proximate analysis, and higher heating value (HHV) measurement are time-consuming. This study developed a closure-aware machine-learning framework for bidirectional cross-prediction between ultimate and proximate properties of thermally treated biomass and for HHV estimation. A literature-derived dataset containing elemental composition, proximate composition, HHV, temperature, and residence time was curated, normalised for compositional closure, and used to train linear, bagging, boosting, and tree-ensemble regression models. Model robustness was evaluated using held-out test sets across five repeated random splits, filtering sensitivity analysis, empirical-formula benchmarking, and HHV circularity ablation. In the ultimate-to-proximate direction, the selected models achieved repeated test-set R 2 values of 0.7555 ± 0.1228 for volatile matter, 0.7018 ± 0.1407 for fixed carbon, 0.6065 ± 0.0905 for ash, and 0.7511 ± 0.1032 for HHV. In the proximate-to-ultimate direction, prediction was less stable, with R 2 values of 0.4246 ± 0.2187 for C, 0.5459 ± 0.2224 for H, 0.5220 ± 0.1182 for N, 0.4131 ± 0.2078 for O, and 0.5707 ± 0.1470 for HHV, whereas S was unreliable. Overall, the proposed framework provides a screening-level tool for cleaner biomass fuel assessment, particularly for major proximate properties and HHV, but not a replacement for laboratory analysis.