Insights into distinct biocarbon assembly and graphitization derived from large-scale bioresource by integrating machine learning with Raman and XRD spectroscopic profiling.

Understanding and predicting the elemental composition and assembly of diverse biocarbon materials are critical for quickly tailoring their physicochemical properties and functional applications. However, it remains challenging to sort out the complex coupling between pyrolysis conditions and carbon structural evolutions. By collecting all major sugars and lignocellulose resources, this study obtained a comprehensive dataset comprising 240 biocarbon samples derived from 17 precursors pyrolysis. Raman spectroscopy and X-ray diffraction profiling were employed to characterize all biocarbon samples, and multiple machine learning models were integrated to predict carbon/C, hydrogen/H, and oxygen/O contents. Among eleven regression algorithms, Gradient Boosting model exhibited the best performance, achieving robust predictions for C, H, and O contents and outperforming linear models in capturing nonlinear structure and composition relationships. Pearson correlation, SHAP interpretation, and partial dependence analysis could consistently sort out the dominant factor of pyrolysis temperature for governing carbon enrichment and H/O depletion, reflecting progressive aromatization, dehydrogenation and deoxygenation during pyrolysis. Structural descriptors including ID/IG, I26°/(I22° + I26°), and interlayer spacing d002, provided complementary information by linking defect density and graphitic ordering to elemental composition. Overall, this work has established an interpretable modeling framework that combines spectroscopic characterization with machine learning to enable rapid and non-destructive prediction of biocarbon composition and property, thereby providing mechanistic insights into biocarbon evolution for specific biocarbon design and assembly.