Designing Machine Learning Interatomic Potentials for Modelling High Entropy Oxides

The recently developed class of novel ceramics known as high entropy oxides (HEOs) are attracting considerable attention from researchers. Due to their large compositional space, they offer a wealth of opportunity for enhanced material design and development. Already they are being investigated as barrier coatings, thermoelectric materials, anodes or solid state electrolytes in batteries, and other applications related to their low thermal conductivity or high ionic conductivity. Computational modelling will allow high throughput screening of this space to predict new stable materials with desirable properties, faster than what is possible in the lab. However, their large compositional space and intrinsic structural disorder correspondingly increases the complexity of calculating their properties computationally. For example, in a randomly placed six-cation rocksalt structure similar to the original HEO, there are of the order 10⁹ ways of placing just the nearest-neighbour cations to a single atom. Complete sampling using ab initio electronic structure methods therefore becomes unfeasibly expensive for these systems. To adequately sample the potential energy landscape of HEOs and calculate their properties computationally, accurate but performant interatomic potentials must therefore be developed. Machine learning interatomic potentials (MLIPs) are uniquely suited to capture the complexity of the many-body interactions involved.<br/><br/>Towards this end, I will present newly parameterized MLIPs for multicomponent and high entropy oxides developed using the ACE and MACE codes, both of which are built around the atomic cluster expansion model. These potentials are capable of achieving density functional theory (DFT) levels of accuracy in the calculation of energies and forces of systems whilst being capable of simulating much larger systems at significantly lower computational cost.<br/><br/>The major challenge facing the development of these potentials is how to build suitable training sets to adequately describe the huge array of interactions in multi component disordered systems. We explore whether universal potentials, trained on millions of structures across the periodic table, are sufficiently accurate and transferable to model the properties and predict the free energies of HEOs. Alternatively, a hyperactive learning approach (HAL) where potentials are trained on a smaller, tailored data set generated itself via a machine learning approach, may be more suited to these configurationally disordered and complex systems.