Computational methods for generating clusters of oxygenated organic molecules

The contribution of oxygenated organic molecules (OOM) in new particle formation is well known yet weakly understood. In this work, we present methods for generating optimized OOM cluster configurations. The framework for configurational sampling, optimizing, and analysing of molecular clusters is provided by JKCS program (Kubečka, Besel, et. al., 2023). Initial sampling of cluster configurations is done at semi-empirical level of theory. Local and global minimum energy configurations are found through successive rounds of filtering a subset of results and re-optimization at higher DFT levels of theory. Finally, we select the lowest energy structures to calculate the binding free energies for each cluster type. 

Experimental research suggests that OOMs with more than 10 carbon atoms contribute to aerosol cluster formation. The size and complexity of OOM clusters significantly limits the number of DFT calculations we can perform, and even very large samplings of cluster configuration space do not guarantee that the global minimum is found. To improve upon existing methods, we introduce constraints to initial sampling which force hydrogen formation between molecules. We also use metadynamics simulations to search for additional local minima. OOMs are observed to cluster in configurations which maximise the number of hydrogen bonds. Thus, the binding free energies are highly dependent on the structure of each molecule and on their ability to form internal hydrogen bonds. OOMs used in this work were obtained using Gecko-AP (Franzon et. al, 2024), a RO2 + RO2 accretion product generator based on the Gecko-A software.

Machine learning force fields have the potential to predict DFT level energies with a fraction of the computational cost. However, most ML force fields do not scale well to larger molecules and fail to correctly model long-range interactions. It is also necessary to sample a dataset which covers the relevant region of the potential energy surface.  We trained a neural network model to predict electronic energies for 2-OOM clusters containing 130-150 atoms. In future work we wish to train a machine learning force field which generalizes to atmospheric molecules and to decrease the prediction error close to chemical accuracy.

References

Kubečka, J., Besel, V., Neefjes, I., Knattrup, Y., Kurtén, T., Vehkamäki, H. and Elm, J. (2023) ACS Omega, 8, 45115. 

Franzon, L., Camredon, M., Valorso, R., Aumont, B. and Kurten T. (2024) Atmos. Chem. Phys., 24, 11679–11699.