Training machine learning potentials with accurate long-range interactions for atmospheric molecular dynamics collision simulations

In the initial stages of atmospheric aerosol particle formation, molecules collide and stick together, forming dimers and small clusters. This is an inherently dynamic process. Molecular dynamics (MD) simulations allow us to model the dynamic behavior of these collision systems. In MD simulations, the trajectory of a system is divided into discrete timesteps, with forces on the nuclei computed at each step to propagate the system. Traditionally, these forces are calculated using classical force fields, which are highly efficient and allow for simulations of large systems over long time scales. However, classical force fields either neglect or only crudely approximate important features like chemical reactions. Atmospheric particle formation depends on proton transfers and cluster reactions to form stable aggregates, making it essential to capture these processes accurately.

Ab initio molecular dynamics (AIMD) use quantum chemistry (QC) calculations to determine the forces on nuclei at each time step. While AIMD can accurately model chemical reactions and other quantum effects, it is computationally unfeasible for anything beyond short simulations of small systems. Recently, machine learning (ML) methods have been applied to create ML potentials for MD simulations. These potentials can replicate high-level QC data while maintaining the efficiency of classical force fields. However, many ML methods rely on a local atomic environment approximation, where the potential is constructed from interactions within a user-defined cutoff radius around each nucleus. This approach fails to capture long-range interactions, which are particularly significant for polar atmospheric molecules like sulfuric acid, as these interactions typically extend well beyond the cutoff radius.

We are training ML potentials for MD simulations of collisions between atmospheric particle-forming molecules, with a focus on accurately capturing long-range interactions. A training set was generated by performing MD simulations of collisions between two sulfuric acid molecules using the semi-empirical GFN1-xTB method, followed by gradient calculations on structures along the collision trajectory. This approach ensures that the free molecules, dimers, and structures along the collision trajectory are well-represented. We then use this dataset to train ML potentials with the paiNN and PhysNet architectures. Both methods rely on the local atomic environment approximation, but PhysNet additionally incorporates long-range electrostatic interactions through learned partial charges and dispersion interactions via Grimme’s D3BJ dispersion correction. By exploring various training sets and model parameters, such as the cutoff radius for the local environment, we aim to develop ML potentials that accurately capture long-range interactions. This project serves as an initial step toward enabling large-scale MD simulations of atmospheric particle formation.