Large-Scale Machine-Learning-Augmented Molecular Dynamics Simulations for Understanding Ice Nucleation on Atmospheric Dust

On Earth, precipitation formation—one of the most important climate processes—is strongly influenced by mineral dust aerosols acting as ice-nucleating particles (INPs). Over the past decade, feldspar minerals have been identified as dominant INPs in the atmosphere, owing to the relatively high freezing temperatures on their surfaces. At the micron scale, experimental studies have demonstrated that surface defects on feldspar are responsible for its remarkable ice-nucleating efficiency. However, the atomic-scale mechanisms underlying ice nucleation on feldspar, and the precise nature of the active sites, remain poorly understood. This project aims to elucidate the atomistic features of feldspar which are responsible for its exceptional ice nucleation ability. To this end, the team will train an efficient and highly-accurate machine learning interatomic potential for the water-feldspar system, able to describe all relevant crystallographic surfaces. The project aims to then apply such a model to drive molecular dynamics simulations and we will carry out a comprehensive investigation of the water structure at these interfaces. Finally, the team will study the ice nucleation process directly using enhanced sampling simulations. In particular, they will perform umbrella sampling calculations to compute the free energy barriers and to elucidate the microscopic mechanism of ice nucleation at feldspar surfaces. The GPU-accelerated computational resources provided by MareNostrum 5 ACC are essential to achieve the ambitious goals described above. This study will have implications for weather prediction and multiple technologies connected to ice formation, such as cloud seeding and artificial snow prediction.