Details zur Publikation |
| Kategorie | Textpublikation |
| Referenztyp | Zeitschriften |
| DOI | 10.1016/j.jece.2026.123911 |
Lizenz  |
|
| Titel (primär) | Modeling adsorption of mobile organic compounds on activated carbon and biochar using machine learning |
| Autor | Saeidi, N.; Vicente, D.J.; Chaudhuri, S.; Georgi, A. |
| Quelle | Journal of Environmental Chemical Engineering |
| Erscheinungsjahr | 2026 |
| Department | TECH |
| Band/Volume | 14 |
| Heft | 5 |
| Seite von | art. 123911 |
| Sprache | englisch |
| Topic | T7 Bioeconomy |
| Supplements | Supplement 1 Supplement 2 Supplement 3 |
| Keywords | Mobile and very mobile organic compounds; Adsorption; Log Kd prediction; Activated carbon; Biochar; Interpretable machine learning |
| Abstract | An interpretable machine-learning model was developed to predict adsorption of mobile and very mobile organic compounds on activated carbon (AC) and biochar (BC) at environmentally relevant trace concentrations (Ce < 5 µg/L). Model development used a harmonized literature dataset of 509 adsorption coefficient values (log Kd) for 74 compounds and was evaluated with independent experimental and literature data (23 log Kd values for 14 compounds), yielding 532 data points. The contribution is a low-concentration, mobile-compound-specific adsorption dataset with descriptors combining solute mobility, pH-dependent charge state, molecular structure, and adsorbent surface properties. The model includes estimated organic carbon/water partition coefficient (log KOC) as a mobility/sorption proxy, pH-dependent charge indicators, aromaticity descriptors, and adsorbent descriptors, including specific surface area, oxygen content, and delta_PZCpH (PZC - pH) as a simple surface-charge descriptor. Adding adsorbent descriptors markedly improved prediction compared with molecular descriptors alone. Random Forest achieved the highest random held-out performance (R2 ≈ 0.81, RMSE ≈ 0.34 across log Kd ≈ 2.3–7.4), whereas compound-grouped cross-validation gave lower performance (median R2 ≈ 0.51, median RMSE ≈ 0.57), indicating more difficult transfer to unseen compounds. Interpretability analyses identified specific surface area, estimated log KOC, molecular size/hydrophobicity-related descriptors, and delta_PZCpH as important contributors, showing that adsorption prediction in the mobile-compound regime cannot be explained by hydrophobicity alone. Overall, the model provides a transparent screening tool for comparing AC/BC adsorption of diverse mobile contaminants within the low-concentration domain, supported by openly available data and code following FAIR principles (https://github.com/NaviddSaeidii/ML-for-mobile-organic-compounds-adsorption-on-AC-BC). |
| Saeidi, N., Vicente, D.J., Chaudhuri, S., Georgi, A. (2026): Modeling adsorption of mobile organic compounds on activated carbon and biochar using machine learning J. Environ. Chem. Eng. 14 (5), art. 123911 10.1016/j.jece.2026.123911 |
|