SCIENCE SPIN-OFFS

[dockzef]

This data-driven paradigm, popularly referred to as machine learning, has been shown by many groups to lead to viable pathways for the creation of interatomic potentials that;(1)surpass conventiona linter atomic potentials both in accuracy and versatility, (2) surpass quantum mechanical methods in cost (by orders of magnitude), and (3) rival quantum mechanics in accuracy,14–16 at least within the con?gurational

This article deals speci?cally with using machine learning methods to create an atomic force prediction capability, i.e., a force ?eld. As recently pointed out, this force ?eld is Adaptive (i.e., new con?gurational environments can be systematically added to improve the versatility of the force ?eld, as required), Generalizable (i.e., the scheme can be extended to any collection of elements for which reliable reference calculations can be performed), and is Neighborhood Informed (i.e., a numerical ?ngerprint that represents the atomic environment around the reference atom is mapped to the atomic force with chemical accuracy).17,18 The force ?eld is henceforth dubbed AGNI.

Learning algorithm
The next vital ingredient required in putting together a predictive framework is the learning algorithm itself. Deep learning neural networks30 and non-linear regression processes15 have been the methods of choice for models describing atomic interactions. Their capability to handle highly non-linearrelations,asisinthecaseof mapping an atom’senvironmenttotheforceitexperiences,

Acknowledgment
This work was supported ?nancially by the Of?ce of Naval Research (Grant No. N00014-14-10098). The authors would like to acknowledge helpful discussions with K. B. Lipkowitz, G. Pilania, T. D. Huan, and A. Mannodi-Kanakkithodi. Partial computational support through a Extreme Science and Engineering Discovery Environment (XSEDE) allocation is also acknowledged.

https://arxiv.org/pdf/1610.02098v1.pdf