(16f) Towards a Deep Learning Potential for Anhydrous Proton Transport.

Standard density functional theory (DFT) molecular dynamic (MD) simulations are prohibitively expensive in terms of compute time and memory requirements when the size of the physical system is larger than several hundred atoms. An alternate approach is to develop a deep learning neural network potential that can accurately reproduce DFT energies within an error of several 10s of meVs. Neural network potentials thus allow large-scale MD simulations for much longer times than can be achieved with DFT-MD, but with comparable accuracy. We have used the recently published DeePMD [1] formalism to develop potentials for systems relevant to modeling proton transport on functionalized graphane. We have used DFT-MD calculations for small structures of graphane and graphanol as training data to build reliable machine-learned potentials. Apart from energies and forces, vibrational and thermal properties such as phonon dispersion relations and thermal expansion coefficients from DFT-MD were used to validate the accuracy of our trained DeePMD potentials. Our goal is to develop a machine learning potential that is able to accurately predict proton conduction on graphanol in the complete absence of water. Previous DFT calculations have shown that graphanol a promising candidate for anhydrous proton exchange membrane applications [2]. We seek to identify large-scale and long-time proton hopping behavior for this system through the use of machine learning potentials.

1. Zhang, L., Han, J., Wang, H., Car, R., & Weinan, E. (2018). Deep potential molecular dynamics: a scalable model with the accuracy of quantum mechanics. Physical Review Letters, 120(14), 143001.
2. Bagusetty, A., & Johnson, J. K. (2019). Unraveling Anhydrous Proton Conduction in Hydroxygraphane. The Journal of Physical Chemistry Letters, 10(3), 518-523.