A new molecular dynamics scheme which combines first-principles molecular dynamics and machine-learning (ML) techniques in a single information-efficient approach coauthored by WCPM member James Kermode has been published in Phys. Rev. Lett this week.
Speaking about the research, Dr Kermode said. “We have been thinking for some time about how to approach optimal information efficiency in these kinds of calculations. The current state of the art approaches are often based on fitting simple models, which risks non-transferability”. This new work shows how the use of ML techniques can help to overcome this problem.
Essentially, the scheme works by going “shopping” in a database of reference configurations every time quantum mechanical (QM) forces are required. If there’s something similar enough, it interpolates within this database using Bayesian inference to predict forces to a high degree of accuracy. If not, a new QM calculation is carried out on the fly to extend the database.
In practical tests in silicon systems, the work shows that progressively fewer calculations are required whenever chemical processes that have been encountered before are seen again.
The scheme is expected to be particularly useful in situations where chemically complex processes are likely to occur in places which are hard to predict beforehand, limiting the applicability of the current non-uniform precision techniques such as quantum mechanics/molecular mechanics (QM/MM). A good example would be the propagation of a crack front in hydrogen embrittlement of steels, which is not yet well understood.
Dr Kermode recently joined WCPM, which has a mission to target the gap between new mathematical and statistical techniques and applications of predictive modelling across science and engineering. The work benefited from an INCITE award for computer time at Argonne Leadership Computing Facility, US, running on the Mira Leadership-class Blue Gene/Q facility.
Z. Li, J. R. Kermode, and A. De Vita, Molecular Dynamics with On-the-Fly Machine Learning of Quantum-Mechanical Forces, Phys. Rev. Lett. 114, 096405 (2015).
For more details, see also a recent WCPM seminar by Dr Kermode (slides available).
Schematic representation of the new “machine learning on the fly” scheme. QM forces on atoms are either predicted as a function of the local atomic environment using Bayesian inference from a database of reference configurations, or computed with an on-the-fly QM calculation and added to the growing ML database.