Development of empirical potentials to replace inaccurate force fields and computationally expensive QM methods is of growing interest in the molecular simulation community. Machine learning (ML) methods are proving to be formidable in their ability to predict energies and forces of molecules. However, development of universal ML potentials, such as those provided by the ANAKIN-ME (ANI) methodology, require vast amounts of QM reference data. For this reason, the parametrization of universal atomistic machine learned potentials can be a daunting task. Data sets generated using random sampling methods, such as the ANI-1 data set, tend to be clustered and sparse in their coverage of chemical space, while also growing too large to be feasible as a training set. Through a newly developed automated active learning algorithm, based on the Query By Committee concept from statistics, we sample chemical space and show a vast improvement of the potential in new areas of chemistry while maintaining a significantly smaller training set. In fact, the new active learned potential is shown to perform exceptionally well in problems involving optimization and molecular dynamics. To directly show that the application of these concepts provides a better description of chemical interactions in a general way, we use this active learning approach to parametrize the ANI potential for Diels-Alder, Claisen and Cope rearrangements. The resulting ANI potential is a chemically accurate and universal potential with a wide range of applications in computational chemistry and physics.<br/>

Active learning in chemical space for the automatic improvement of the ANI deep learned potential with an application to reaction profiles

255th National Meeting (2018)

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.