Addressing uncertainty in atomistic machine learning

Andrew A. Peterson, Rune Christensen, Alireza Khorshidi

Research output: Contribution to journalJournal articleResearchpeer-review

721 Downloads (Pure)

Abstract

Machine-learning regression has been demonstrated to precisely emulate the potential energy and forces that are output from more expensive electronic-structure calculations. However, to predict new regions of the potential energy surface, an assessment must be made of the credibility of the predictions. In this perspective, we address the types of errors that might arise in atomistic machine learning, the unique aspects of atomistic simulations that make machine-learning challenging, and highlight how uncertainty analysis can be used to assess the validity of machine-learning predictions. We suggest this will allow researchers to more fully use machine learning for the routine acceleration of large, high-accuracy, or extended-time simulations. In our demonstrations, we use a bootstrap ensemble of neural network-based calculators, and show that the width of the ensemble can provide an estimate of the uncertainty when the width is comparable to that in the training data. Intriguingly, we also show that the uncertainty can be localized to specific atoms in the simulation, which may offer hints for the generation of training data to strategically improve the machine-learned representation.
Original languageEnglish
JournalPhysical Chemistry Chemical Physics
Volume2017
Issue number18
Pages (from-to)10978-10985
ISSN1463-9076
DOIs
Publication statusPublished - 2017

Fingerprint

Dive into the research topics of 'Addressing uncertainty in atomistic machine learning'. Together they form a unique fingerprint.

Cite this