Rapid Fourier-transform infrared spectroscopic profiling of generation and graphitization of biocarbon from large-scale bioresource by integrating optimal machine learning

The physicochemical properties of biocarbons—carbon-rich materials derived from biomass pyrolysis—directly determine their application in energy conversion, environmental remediation, and agricultural improvement. However, conventional characterization methods typically rely on high-precision instruments that are costly and complicated to operate, thereby limiting the rapid assessment of diverse biocarbon materials. Herein, 112 biocarbon samples were prepared from diverse biomass resources, including all major sugars and representative lignocellulosic substrates, via standard pyrolysis at 300–600 °C. The Fourier-transform infrared (FTIR) data were then used to develop six machine learning (ML) models to simultaneously predict the elemental content (C, H, O) and graphitization parameters ( I D / I G ratio) of all biocarbon samples. The AdaBoost model demonstrated the best performance in predicting the C and O content, with coefficients of determination ( R 2 ) of 0.886–0.902 for the test set. In comparison, the two-layer artificial neural network achieved the highest accuracy for the H content and I D / I G ratio, with R 2 values of 0.932 and 0.901, respectively. SHapley Additive exPlanations (SHAP) interpretability analysis further identified the pyrolysis temperature as the most critical factor influencing the composition and graphitization of biocarbon. Therefore, this study established a high-throughput, high-accuracy methodology for characterizing distinct biocarbon substrates via direct FTIR screening, enabling reliable and rapid assessment of the development of specific properties and the functional application of green carbons.