A dominance-based ternary framework for ML-guided interpretable analysis of biomass pyrolysis products

Biomass pyrolysis is a complex thermochemical process in which bio-oil, biochar, and gas are simultaneously generated under coupled material and energy constraints. Although machine learning (ML) has been widely applied to predict individual product yields with high accuracy, most existing studies primarily emphasize statistical performance and provide limited insight into the structured relationships governing multi-product distribution. This study introduces a dominance-based ternary regime mapping method that converts ML-predicted yields into interpretable oil/char/gas dominance regions under controlled compositional scanning. A data-driven yet physically interpretable analytical framework is proposed to investigate biomass pyrolysis behaviour by combining controlled compositional data scanning with a dominance-based ternary representation. A comprehensive dataset comprising over 1,400 experimental records from more than 100 studies was compiled, covering biomass proximate and ultimate analyses, lignocellulosic composition, and pyrolysis operating conditions. To ensure data consistency, the dataset was harmonized through basis standardization and mass-closure filtering. Proximate analysis variables (volatile matter, fixed carbon, and ash) and elemental composition variables (C, H, and O) were systematically scanned within the min-max bounds observed in the dataset, while other variables were fixed. A dominance-based ternary framework was then employed to map experimental data into a unified compositional space, explicitly distinguishing oil-rich, char-rich, gas-rich, and mixed regimes. The results show that temperature governs product dominance migration, while feedstock compositions impose systematic, yet non-deterministic, geometric constraints. This framework enables investigation of competitive multi-product formation pathways in biomass pyrolysis.