WEB Active learning of surrogate interatomic potentials from large-scale simulations with application to dislocation motion in tungstenFriday (25.09.2020) 09:00 - 09:15 M: Modelling and Simulation 2 Part of:
One of the main open challenges in computational materials science is the construction of interatomic potentials which achieve quantitative agreement with fully—but infeasible—ab initio models for large-scale problems, possibly involving tens of thousands of atoms. Empirical potentials, arguably the most popular type of interatomic potentials, generally fail in making quantitative predictions. Therefore, they largely remain inadequate for a predictive modeling of multicomponent systems and, thus, for a computational discovery of new materials.
The advent of machine learning interatomic potentials (MLIPs) for small systems of order 10–100 yet holds promise that overcoming these limitations appear within sight (see  for a review). On the other hand, a well-known drawback of state-of-the-art MLIPs is their poor ability to extrapolate beyond the training domain. This demands for carefully chosen, problem-dependent training data which can hardly be defined by the user prior to a simulation due to the vast amount of possible atomic neighborhoods to be considered. Therefore, solution algorithms for atomistic problems using MLIPs have to incorporate error estimators which measure the degree of extrapolation in order to "find" the right training data during a simulation.
Here, we present such an algorithm based on active learning  and apply it to screw dislocation motion in bcc tungsten . Using structural identification methods, we locate the dislocation line, check the extrapolation grade of the MLIP around its core, and construct new training configurations, sufficiently small to be computable with plane-wave density functional theory, if the extrapolation grade exceeds some threshold. We show that the algorithm reproduces classical results from the literature (i.e., core structure, Peierls barrier/stress), in addition to its anticipated application to more complex problems involving interactions of dislocations with other types of defects, e.g., interstitials, which are out of scope with ab initio methods solely.
 Y. Zuo et al., 2019. Performance and Cost Assessment of Machine Learning Interatomic Potentials. J. Phys. Chem. A 124, 731–745
 E. Podryabinkin, A. Shapeev, 2017. Active learning of linearly parametrized interatomic potentials. Comput. Mater. Sci. 140, 171–180
 M. Hodapp, A. Shapeev, 2020. In operando active learning of interatomic interaction during large-scale simulations. arXiv: 2004.13158