Molecular dynamics with on-the-fly machine learning of quantum mechanical forces

Zhenwei Li, James Kermode (King's College London, London, UK), and Alessandro De Vita

We present a molecular dynamics scheme where forces on atoms are either predicted by a machine learning (ML) scheme or, if necessary, computed with on-the-fly quantum mechanical (QM) calculations and added to the growing ML database [1]. The resulting force field is accurate and transferable, since QM accuracy can be enforced to any desired tolerance and database completeness is never required. The scheme is also efficient since the frequency of QM calls systematically decreases, eventually falling to zero in simple situations where no novel chemical processes or bonding geometries are encountered during the simulation.

We expect the method to be particularly useful for the simulation of processes where complex but recurring chemical steps are encountered (e.g. crack propagation [2]), which can be learned, while time-localised occurrences of new chemical bonding geometries (e.g. crack-impurity interactions [3] or reactions with environmental molecules [4]) cannot be ruled out, so that a fixed classical potential is not an option.

[1] Z. Li, J. R. Kermode and A. De Vita, Submitted (2014).
[2] J. R. Kermode, T. Albaret, D. Sherman, N. Bernstein, P. Gumbsch, M. C. Payne, G. Csányi, and A. De Vita, Nature 455, 1224 (2008).
[3] J. R. Kermode, L. Ben-Bashat, F. Atrash, J. J. Cilliers, D. Sherman, and A. De Vita, Nat. Commun. 4, 2441 (2013).
[4] A. Gleizer, G. Peralta, J. R. Kermode, A. De Vita, and D. Sherman, Phys. Rev. Lett. 112, 115501 (2014).