In many ways, advances in machine-learning have transformed computational chemistry. One such leap forward is the advent of machine-learning interatomic potentials (=force-fields). Unlike traditional classical potentials, which rely on predefined functional forms, machine-learned interatomic potentials (MLIPs) learn directly from quantum-chemical data, typically DFT. The parametrization generally requires large, high-quality datasets, and the transferability of the force-field remains a challenge. Fortunately, we have acquired some considerable experience in the field of MLIPs in recent years. This proposal concerns applications of MLIPs. The systems in focus are water structures "on top of" ceria surfaces. Cerium oxide ("ceria", CeO2) has intriguing physical and chemical properties and is an essential part of many high-tech devices. In many of these applications water is present, by design or as an unwanted ingredient. The objective of our planned simulations is to find the characteristics of the interactions and structures present at the interface between water and ceria for three experimentally found ceria surfaces with different structures. For this we will primarily use MLIP-based MD simulations and, of course, adequate post-processing. One goal is to be able to perform droplet calculations, i.e. deposit water on the surfaces and examine the water spreading tendency on the surface, as nano-ceria has been found experimentally to be water-repellent, a property that is technically exploitable. Such simulations require very long and large-size (the droplets must not touch initially) simulations, which would be unfeasible using ab initio MD simulations. The calculations planned for this project are as follows. 1. Perform static and dynamic periodic water/solid DFT calculations (using CP2K) to fill the "holes" in our existing training datasets. We already have a reasonably extensive dataset to start with as we have performed many smaller-scale MD simulations in the past and collected trajectories. 2. Training step: The MLIP training will be performed on GPUs using local resources. 3. Testing on Dardel. MLIPs will be tested using small-scale MD simulations, primarily assessing trajectory stability (e.g. explosions and energy drift). Snap-shots from failed runs will augment the training set in an active learning loop. Once stable trajectories are obtained, results will be compared to the DFT MD results to determine whether to add data or proceed to step. 4. Perform water-droplet-on-surface production MD simulations with the new MLIP and monitor the interface structures and water spreading/confinement. The role of molecular dissociation of the first (surface-bound) water layer for the adhesion of the rest of the water will be examined, as well as surface roughness. 5. Perform static DFT calculations to examine some of the most valuable electronic properties such as charge distributions, which get lost in the MLIP approach.