Machine learning-enabled characterization of coal thermal dissolution products by FTIR and mass spectrometry with ionization bias calibrations

Accurate molecular-level characterization of coal thermal dissolution (TD) products is hindered by structural complexity and mass spectrometry (MS) ionization bias. This study combines Fourier transform infrared (FTIR) spectroscopy and Orbitrap MS in a sample-identity-based framework that links fraction-level functional-group descriptors with formula-level MS records. To reduce ionization-dependent distortions in MS-derived relative abundances, an elemental-analysis-guided ionization-bias correction was applied, yielding bias-adjusted semi-quantitative molecular-class distributions. Comparative analysis of 48 fractions from 12 coals revealed a distinct ionization divide: electrospray ionization favors polar N3–N4 species, while atmospheric pressure chemical ionization extends coverage to lower-polarity O1–O2 and N1 compounds. A supervised machine-learning framework was used to classify 13 molecular families and to compare FTIR-only, MS-only, and fused FTIR–MS feature configurations. The MS-only configuration yielded higher grouped-cross-validation macro-F1 than the fused configuration in all four classification tasks, whereas external evaluation showed task-dependent changes after addition of the fraction-level FTIR descriptors. For the grouped-CV-selected MS-only models, pooled external macro-F1 ranged from 0.568 to 0.833, balanced accuracy from 0.702 to 0.840, and Cohen’s Kappa from 0.505 to 0.774 across the four classification tasks. Cross-modal co-distribution patterns were consistent with an association between oxygen-containing functional-group features and nitrogen-containing molecular classes during TD. The combined analytical workflow provides a framework for comparing fraction-level functional-group information with formula-level molecular composition in coal TD products.