Details zur Publikation |
| Kategorie | Textpublikation |
| Referenztyp | Zeitschriften |
| DOI | 10.1016/j.gca.2026.08.034 |
Lizenz  |
|
| Titel (primär) | Structure-informed prediction of organic matter thermodynamics from quantum chemistry and machine learning |
| Autor | Sibille, T.P.; Knorr, K.-H.; Lechtenfeld, O.J.
; Lau, M.P. |
| Quelle | Geochimica et Cosmochimica Acta |
| Erscheinungsjahr | 2026 |
| Department | EAC |
| Band/Volume | 431 |
| Seite von | 247 |
| Seite bis | 259 |
| Sprache | englisch |
| Topic | T9 Healthy Planet |
| Daten-/Softwarelinks | https://doi.org/10.5281/zenodo.22025508 |
| Supplements | Supplement 1 |
| Keywords | Dissolved organic matter (DOM); Gibbs energy of oxidation half reaction per carbon (ΔG° (Cox)); Nominal oxidation state of carbon (NOSC); Density functional theory (DFT) |
| Abstract | Thermodynamic constraints shape the degradation and persistence of dissolved organic matter (DOM) in aquatic environments. Predicting the Gibbs energy of the oxidation half reaction per mole of carbon (ΔG°(Cox)) is therefore central to assessing the thermodynamic potential of organic matter degradation. A widely used approach is the linear relationship between nominal oxidation state of carbon (NOSC) and ΔG°(Cox), introduced by LaRowe and Van Cappellen, which provides a convenient and versatile single-proxy estimate. However, NOSC is a stoichiometric descriptor and assigns identical ΔG°(Cox) values to formulae that reflect structures differing by more than 90 kJ molC-1. This underlying diversity of structures can be isomeric, but can also reflect variation in molecular size, heteroatom composition, hydrogen abundance, and level of unsaturation. Here we present a structure-informed Random Forest model for ΔG°(Cox) trained on a curated set of over 3,500 quantum-chemically calculated DOM-relevant molecules spanning diverse elemental compositions and structures. The model was trained using formula-derived descriptors, including elemental ratios, molecular mass (Mr), double bond equivalence (DBE), heteroatom content, NOSC and others. Predictions are made at the molecular level and represent an average across potential isomers of a formula. It relies on ΔG°(Cox) data derived with the Becke three-parameter Lee-Yang-Parr density functional (B3LYP) and captures thermodynamic variability that the NOSC-based linear relationship systematically misses. As proof of concept, the model was applied to ultra-high resolution mass spectrometry (UHR-MS) data for four benchmark DOM reference materials, yielding ΔG°(Cox) distributions that are broader and multimodal compared to the NOSC-based estimates. These results demonstrate that machine learning models grounded in quantum chemical data provide a more differentiated thermodynamic representation of DOM formula space and support improved prediction of organic matter oxidation potential. |
| Sibille, T.P., Knorr, K.-H., Lechtenfeld, O.J., Lau, M.P. (2026): Structure-informed prediction of organic matter thermodynamics from quantum chemistry and machine learning Geochim. Cosmochim. Acta 431 , 247 - 259 10.1016/j.gca.2026.08.034 |
|