University of Cambridge > > Machine learning in Physics, Chemistry and Materials discussion group (MLDG) > Machine learning an interatomic potential without (much) human effort

Machine learning an interatomic potential without (much) human effort

Add to your list(s) Download to your calendar using vCal

If you have a question about this talk, please contact Bingqing Cheng .

While many of the applications of machine learning (ML) methods in materials have been to the direct prediction of experimentally observable properties, the same methods for regression in high dimensions have also been very successful in approximating high accuracy, e.g. density functional theory (DFT), potential energy surfaces. Such an approximation constitutes an interatomic potential: an explicit formula for the bonding energy as a function of the atomic positions. Since the input space is large and the effective functional form of ML methods is extremely flexible, it is easy to inadvertently produce a fit that is accurate near the input data, but extremely inaccurate for reasonable configurations that are just a bit different. Since such a fit leads to artifacts in any real simulation, it has proven essential to develop an extensive fitting database that includes not only all the relevant configurations but also the boundary of the important, i.e. relatively low energy, region. Doing this ensures that the resulting potential has an increasing energy as it leaves the low energy configuration subspace. Our experience in developing several such fitting databases has shown that doing this manually is a time consuming procedure that would greatly benefit from automation and broadly applicable heuristics. We have developed such a process, which iterates over a sequence of random- structure searches and potential fitting, with only a minimal number of reference (DFT) evaluations. The resulting potentials are accurate and robust for the configurations that occur during the random structure search, from the high-energy regions all the way to local minima; they also give at least qualitatively reasonable values for defects such as vacancies and surfaces. We apply the process to a range of materials with different chemical nature and coordination environments, including insulating, semiconducting, and metallic bonding.

This talk is part of the Machine learning in Physics, Chemistry and Materials discussion group (MLDG) series.

Tell a friend about this talk:

This talk is included in these lists:

Note that ex-directory lists are not shown.


© 2006-2020, University of Cambridge. Contact Us | Help and Documentation | Privacy and Publicity