Machine-learned potentials for eucryptite: A systematic comparison

JR Hill and W Mannstadt, JOURNAL OF MATERIALS RESEARCH (2023).

DOI: 10.1557/s43578-023-01183-7

Three machine-learned potentials (SNAP, NNP, ACE) were created from the same training set of DFT energies and forces for a total of 1024 structures. DFT calculations were performed with the PBE functional and the Grimme D3 corrections. DFT energies can be reproduced within a few meV by the potentials. The potentials are evaluated how they predict structures, thermal expansion coefficients, and ionic conductivities of alpha- and beta-eucryptite. Structures and thermal expansion coefficients are in good agreement with experimental values. All potentials reproduce the negative thermal expansion coefficient along the c axis of beta-eucryptite, although only ACE calculates a negative thermal expansion coefficient for the volume. Ionic conductivities can be predicted only qualitatively correct. Molecular dynamics simulations performed with some of the potentials at higher temperatures can result in unphysical structures.

Return to Publications page