The aim of this project is to investigate how the recently developed Allegro [1] equivariant graph neural network (GNN) interatomic potential performs in predicting phase transformations and diffusion in a complex, alloyed, crystalline solid. Machine learned interatomic potentials (MLIPs) have shown great promise in replacing Density Functional Theroy (DFT) based ab inito molecular dynamics (AIMD) simulations. Reaching comparable levels of accuracy at a much reduced computational cost. It has long been realised that having the MLIP respect the underlying physical symmetries of the system is of great importance for the accuracy and stability of the MLIPs. Conventionally, this has been achieved by describing the geometry of the system in terms of symmetry respecting descriptors based on distances and angles. In the last few years, several equivariant GNN MLIPs have emerged, which acts directly on the raw atomic position vectors, but respect symmetries by using only so symmetry respecting internal operations in the NN. While these equivariant MLIPs have been extensively evaluated, and shown highly promising results, on well-known "standard" data sets, examples of their application to complex, "real-world", materials are still somewhat rare. Here, we propose to train and apply Allegro GNNs for a set of alloyed Na-ion conductors, recently investigated for their potential use as solid-state electrolytes in Na-ion batteries [3] . Using a by us recently generated pre-existing DFT-based dataset of energies, forces and stress tensors for a range of chemical compositions and atomic configurations, we will train equivariant GNNs, aiming to describe the sodium thiophosphate ion conductors with the empirical formula Na_(3-x)Sb_xW_(1-x)S_3, throughout the range of x= 0 to x=0.25. This will require the trained GNN to describe (1) the diverse local chemical environments of an alloy, (2) ionic diffusion and (3) structural phase-transformations, a highly demanding set of tasks. A central aim is to investigate the data efficiency of the Allegro GNN. In particular, we set out to answer the question: What chemical, and structural diversity needs to be present in the training set for the Allegro GNN to accurately describe these complex systems and phenomena? [1] https://github.com/mir-group/allegro; Nature Comms. 2023 ( https://www.nature.com/articles/s41467-023-36329-y ) [2] Nature Comms. (2022) ( https://www.nature.com/articles/s41467-022-29939-5 ) [2]J. Am. Chem. Soc. (2023); ( https://pubs.acs.org/doi/full/10.1021/jacs.2c11803 )