Mesoscale models from massively parallel atomistic simulations: uncertainty driven, self-optimizing strategies for hard materials – MeMoPAS

Atomistic simulations utilising short ranged interatomic force fields can be parallelized in space but not in time, limiting trajectory timescales to microseconds irrespective of the number of available processors. As material evolution is routinely controlled by rare events over the timescales of milliseconds to years it is therefore essential to coarse grain atomistic simulations into mesoscale models to allow access to the time and length scales of significance for materials science. Error in this process arises from two main sources.

First, a priori or intuition based coarse graining can have potentially disastrous assumptions on the possible underlying atomistic processes. As a result, the produced mesoscale model is fundamentally unable to capture the behaviour of the true atomistic system, meaning a more agnostic approach must be developed.
Second, implicit electron force fields are always models of the Hellman-Feynman forces driving Born-Oppenheimer dynamics. Whilst traditional empirical forms are able to accurately capture certain regions of a material's energy landscape, force field development is being revolutionised by the non-linear regression techniques of machine learning, which are increasingly capable of ab initio accuracy. Nevertheless, such accuracy can only be achieved if the training database of configurations to which the regression is applied is sufficiently diverse and relevant for the problem under study.

To solve these problems at the petascale and exascale, autonomous uncertainty quantification is of central importance. This research program will develop Bayesian estimators of atomistic sampling completeness to optimize the automated, massively parallel construction of mesoscale models. The discovered mechanisms and uncertainty quantifiers will also be used to improve force field parametrization in kinetically important regions of configuration space. The goal is an exascale-efficient, autonomous, uncertainty-aware coarse graining scheme that allows quantitative insight into complex mechanisms of material evolution.