Over the last few decades, computational chemists, drug developers, and materials scientists have been dependent on computationally efficiency force fields for large scale simulation. While force fields have revolutionized these studies, they tend to be less accurate and transferability than computationally expensive quantum mechanical (QM) methods. Highly efficient machine learning-based (ML) methods have been shown to accurately learn QM calculated potential energies and forces of atomistic systems. The success of these ML potentials has been driven largely by the development of groundbreaking methods of data selection such as active learning. Active learning methods automatically maximize data diversity by selecting new data where the model predicts bad performance. Data diversity is shown to produce the most transferable models, but also unearths some difficulties in learning certain physical phenomena.  In this work we compare different methods and philosophies of data selection for building ML potentials, including what techniques lead to the most overall transferable models and best accuracy.

Importance of data selection for machine learning-based atomistic potentials

2019 Fall National Meeting

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