467637 Machine-Learning Acceleration of the Exploration of Potential Energy Surface within Amp

Friday, November 18, 2016: 2:48 PM
Yosemite A (Hilton San Francisco Union Square)
Alireza Khorshidi, Brown University, Providence, RI and Andrew A. Peterson, Brown University

Novel nanostructures with enhanced properties are required to be computationally designed in various fields of science and engineering, e.g. nanomaterials with improved load-bearing properties in materials science [1] or catalysts to speed up reaction kinetics in chemistry [2]. In principle, accurate calculations can be carried out based on electronic structure methods. However, with the current status of computational facilities, such accurate calculations are feasible only on relatively small systems within limited time scale. In other words the accuracy might have to be sacrificed in favor of less computational expense as the length and time scales of simulations increase. As the result, emergence of new computational methods seems essential.

Pattern recognition of energetics from an ab initio database can help reducing the number of required quantum mechanical calculations, as the result making larger-length or longer-time scale simulations practical. This can be done inside our open-source modular Atomistic Machine-learning Package (Amp) [3], which has been recently developed by the authors of this paper. Amp has two main components, a descriptor for modeling local atomic environments and a regression model to establish the relationship between energies/forces with local atomic environment. Amp is developed compatible with the widely-used open-source Atomic Simulation Environment (ASE), facilitating interactive cooperation between the Amp machine-learning calculator and any ab initio calculator within ASE. This will be illustrated here in the context of two well-known problems in molecular sciences, a) determination of the structure of stable nanoclusters, and b) finding the minimum energy pathway of reactions. It will be shown that machine-learning can remarkably improve the exploration of the potential energy surface.

References

[1] Q. Qin et al.; Recoverable plasticity in penta-twinned metallic nanowires governed by dislocation nucleation and retraction, Nature Communication, 6, 2015.

[2] S. Freakley et al.; Palladium-tin catalysts for the direct synthesis of H2O2 with high selectivity, Science, 351, 2016.

[3] A. Khorshidi, A. A. Peterson; Amp: A modular approach to machine learning in atomistic simulations, Computer Physics Communications, 2016.


Extended Abstract: File Not Uploaded