A1 Refereed original research article in a scientific journal
Predicting gas-particle partitioning coefficients of atmospheric molecules with machine learning
Authors: Lumiaro Emma, Todorovic Milica, Kurten Theo, Vehkamäki Hanna, Rinke Patrick
Publisher: COPERNICUS GESELLSCHAFT MBH
Publication year: 2021
Journal: Atmospheric Chemistry and Physics
Journal name in source: ATMOSPHERIC CHEMISTRY AND PHYSICS
Journal acronym: ATMOS CHEM PHYS
Volume: 21
Issue: 17
First page : 13227
Last page: 13246
Number of pages: 20
ISSN: 1680-7316
eISSN: 1680-7324
DOI: https://doi.org/10.5194/acp-21-13227-2021
Self-archived copy’s web address: https://research.utu.fi/converis/portal/detail/Publication/67222683
The formation, properties, and lifetime of secondary organic aerosols in the atmosphere are largely determined by gas-particle partitioning coefficients of the participating organic vapours. Since these coefficients are often difficult to measure and to compute, we developed a machine learning model to predict them given molecular structure as input. Our data-driven approach is based on the dataset by Wang et al. (2017), who computed the partitioning coefficients and saturation vapour pressures of 3414 atmospheric oxidation products from the Master Chemical Mechanism using the COSMOtherm programme. We trained a kernel ridge regression (KRR) machine learning model on the saturation vapour pressure (P-sat) and on two equilibrium partitioning coefficients: between a water-insoluble organic matter phase and the gas phase (K-WIOM/G) and between an infinitely dilute solution with pure water and the gas phase (K-W/G). For the input representation of the atomic structure of each organic molecule to the machine, we tested different descriptors. We find that the many-body tensor representation (MBTR) works best for our application, but the topological fingerprint (TopFP) approach is almost as good and computationally cheaper to evaluate. Our best machine learning model (KRR with a Gaussian kernel + MBTR) predicts P-sat and K-WIOM/G to within 0.3 logarithmic units and K-W/G to within 0.4 logarithmic units of the original COSMOtherm calculations. This is equal to or better than the typical accuracy of COSMOtherm predictions compared to experimental data (where available). We then applied our machine learning model to a dataset of 35 383 molecules that we generated based on a carbon-10 backbone functionalized with zero to six carboxyl, carbonyl, or hydroxyl groups to evaluate its performance for polyfunctional compounds with potentially low P-sat. The resulting saturation vapour pressure and partitioning coefficient distributions were physico-chemically reasonable, for example, in terms of the average effects of the addition of single functional groups. The volatility predictions for the most highly oxidized compounds were in qualitative agreement with experimentally inferred volatilities of, for example, alpha-pinene oxidation products with as yet unknown structures but similar elemental compositions.
Downloadable publication This is an electronic reprint of the original article. |