Development of Machine Learning Interatomic Potentials to model corrosion behavior in Molten Salt Reactors

In this study, we developed and evaluated the efficacy of Machine Learning Interatomic Potentials (MLIPS) designed for a molten salt/high entropy alloy interface at extreme temperatures and its relevance toward corrosion behavior. We implemented both ML algorithms and neural networks, ranging from Moment Tensor Potentials (MTP) to Equivariant and Invariant deep learning architectures, to study the scalability and accuracy of different training methodologies. To capture the cluster dynamics and the initial corrosion mechanism at the interface we created many complex training configurations and optimized the hyperparameters for each model. Using molecular dynamic (MD) simulations we analyze the charge clustering of the molten salt cations at the high entropy alloy surface, and we compare our results with experimental observations. Furthermore, we show that these potentials have achieved chemical accuracy, and that the system is chemically stable at extreme conditions similar to those inside of a Molten Salt Reactor (MSR). We acknowledge the support from the NASA Space Grant Summer Fellowship and the DOE's access to Perlmutter HPC workstation.