The past decade has seen Machine Learning emerge as a prominent new tool in molecular modelling. In particular, Machine Learning Interatomic Potentials (MLIPs) are methods that can reproduce potential energy surfaces with ab initio accuracy at a fraction of the cost of DFT and Quantum Chemistry methods. The resulting lower cost of the MLIP models enables the simulation of larger systems (100x) and longer simulation times scales (100x) to be routinely studied. This promises to revolutionize research sectors that benefit from molecular simulations such as novel material discovery (energy), drug design and protein modelling (medical). In the recent years, the field has seen multiple developments taking place. While earlier MLIPs were based on Bayesian statistics (e.g. linear models, kernel models), recent advances in computing (e.g. powerful GPUs) led to the development of neural network models (NN). These models treat data in a very high-dimensional space, making use of millions of parameters, which allows them to capture complex patterns and interpolate on massive amounts of data. The critical process behind NN-MLIP effectiveness is training which involves feeding the network different configurations labelled with DFT evaluated properties (energy, forces etc.) and tuning the network's internal parameters until it predicts these properties accurately. Training generally requires non-trivial and meticulous selection of data and the resulting models often display poor transferability to systems they have not been trained on. This project aims to improve our understanding in this area by exploring the capabilities of newly developed MACE foundational MLIPs (FMLIPs) for modelling complex battery materials and interfaces - a particularly challenging and interesting system to study. MACE is a message-passing graph-neural network MLIP that has proven very effective in learning the chemical space and producing stable molecular dynamics in a wide variety of systems ranging from inorganic crystals and molecular liquids to complex organic-inorganic interfaces. The MACE foundational models have been pretrained on massive datasets containing various systems and the hypothesis (which will be thoroughly tested in this PhD project) is that they can be used to model novel systems with little retraining or fine-tuning. If proven true, FMLIPs have the potential to dramatically reduce the amount and complexity of training, making models even cheaper and accessible to a wider scientific audience. This project will investigate how fine-tuning can be optimized to ensure models are stable in MD, and accurate not only on reproducing energies and forces but also the thermodynamics and kinetics of the systems under study. This project will focus on a particularly challenging system for FMLIPs: battery materials and interfaces. Specifically, the electrolyte mixture of Li-Ion batteries will be studied with the aim to improve our understanding of battery properties such as charging speed and lifetime which are underpinned by atomic-level processes. The electrolyte mixture is a crucial component of batteries that facilitates ion transport between electrodes. The long timescales required to capture important properties (10-100ns) and the complexity of the molecular liquid make the electrolytes a difficult system to simulate. Special-purpose MACE MLIPs have recently been shown to produce stable dynamics for the desired timescales and accurately predict properties both within and between molecules of the electrolyte. This work will explore whether fine-tuning general-purpose MACE FMLIPs for electrolytes can yield the same results with significantly less human effort. If true, this will open the possibility to easily retrain on different levels of ab initio theory and compare first-principle predictions directly to experiment.