Solvents - with water just being one prominent example - critically determine the stability and properties of chemical species in the liquid phase. As such, they govern the atomistic mechanisms of chemical reactions not only in the bulk liquid but also in ubiquitous processes at interfaces, from solid surfaces of metals and oxides to biological systems. Prominent examples are electrochemistry, where electrodes are in contact with an electrolyte, and large biomolecules or their aggregates, such as lipid membranes and proteins, which cannot take their role from self-assembly to protein folding without the presence of a surrounding solvent as one of the key parameters affecting these processes. Hence, a robust atomic scale description of liquids and their interfaces, including dynamics and kinetics, are urgently required for knowledge-driven advances in science and technology.

Consequently, the theoretical study of such systems via computer simulations is a very active field and has tremendously contributed to deepening our knowledge in these topics. Nevertheless, advances in computer simulations are still hampered by the challenge to describe many different types of interactions, from covalent bonds via ionic and dispersion interactions to metallic bonding, at the same level of accuracy. While electronic structure methods can in principle provide a reliable description of these systems, their high computational costs limit their applicability to large numbers of atoms and long time scales. On the other hand, classical potentials, while efficient, often only provide limited insights.

In recent years, Machine Learning Potentials (MLPs) have attracted a lot of attention as they promise a solution to these challenges, and many different methods have been proposed to date. By training on datasets derived from accurate quantum mechanical calculations, MLPs can in principle approximate the complex multidimensional potential energy surface with ab initio accuracy while yielding energies and forces at drastically reduced computational costs. There are now many examples of the application of MLPs to liquids like water, molecules, and materials like metals and oxides. However, MLPs for the combination of these subsystems are still in their infancy, and only a few simulations of complex molecules in solution and of ideal single-crystal solid-liquid interfaces have been reported. For significant advances, numerous open challenges need to be overcome, which prevent the routine application of MLPs to complex systems. These are

• Access to accurate and robust reference data, in particular beyond DFT quality
• Improving the accuracy, transferability, and speed of MLPs
• Efficient large-scale simulations with accurate electrostatics employing variable charges
• More accurate incorporation of non-bonded long-range interactions
• Improved training strategies for interfaces and molecular solutions covering a large configuration space

Numerous groups worldwide contribute to advancing these challenges, with a plethora of strategies and methodologies. However, no single group will be able to resolve the open challenges. Improved coordination between different researchers is required for faster progress in the field.