Description
Molecular dynamics (MD) is a powerful condensed matter simulation tool for bridging between macroscopic continuum models and quantum models (QM) treating a few hundred atoms, but it is limited by the accuracy of the interatomic potential. Sound physical and chemical understandings of these interactions have resulted in a variety of concise potentials for certain systems, but it is difficult to extend them to new materials and properties. The solution is obvious but challenging: develop more complex potentials that reproduce large QM datasets. In this discussion I will discuss two different ways that we are pursuing this goal. The first approach uses the ReaxFF family of potentials that reproduces known chemical reaction pathways in small clusters of atoms while still allowing molecular dynamics simulations of millions of atoms undergoing chemical reaction. The second approach, SNAP, is a very general machine-learning approach for automated generation of interatomic potentials from large QM datasets. I. Initiation in energetic materials is fundamentally dependent on the interaction between a host of complex chemical and mechanical processes, occurring on scales ranging from intramolecular vibrations through molecular crystal plasticity up to hydrodynamic phenomena at the mesoscale. A variety of methods [e.g. quantum electronic structure methods (QM), nonreactive classical molecular dynamics (MD), mesoscopic continuum mechanics] exist to study processes occurring on each of these scales in isolation, but cannot describe how these processes interact with each other. In contrast, the ReaxFF reactive force field, implemented in the LAMMPS parallel MD code, allows us to routinely perform multimillion-atom reactive MD simulations of shock-induced initiation in a variety of energetic materials. II. The growing availability of large QM data sets has made it possible to use automated machine-learning approaches for interatomic potential development. Bartok et al. demonstrated that the bispectrum of the local neighbor density provides good regression surrogates for QM models using Gaussian process regression. We adopted a similar bispectrum representation within a linear regression scheme that we called SNAP. We have produced potentials for tantalum and indium phosphide. Results will be presented demonstrating the accuracy of these potentials relative to the training data, as well as their ability to accurately predict material properties not explicitly included in the training data. Comparing to recent QM calculations of screw dislocation cores in BCC tantalum, we observe that the SNAP potential gives the correct core structure and the correct energy barrier for screw dislocation motion, unlike existing EAM and ADP potentials.
Recommended Citation
Thompson, A. (2014). Large-scale atomistic materials simulation using quantum-accurate interatomic potentials. In A. Bajaj, P. Zavattieri, M. Koslowski, & T. Siegmund (Eds.). Proceedings of the Society of Engineering Science 51st Annual Technical Meeting, October 1-3, 2014 , West Lafayette: Purdue University Libraries Scholarly Publishing Services, 2014. https://docs.lib.purdue.edu/ses2014/mms/imms/7
Large-scale atomistic materials simulation using quantum-accurate interatomic potentials
Molecular dynamics (MD) is a powerful condensed matter simulation tool for bridging between macroscopic continuum models and quantum models (QM) treating a few hundred atoms, but it is limited by the accuracy of the interatomic potential. Sound physical and chemical understandings of these interactions have resulted in a variety of concise potentials for certain systems, but it is difficult to extend them to new materials and properties. The solution is obvious but challenging: develop more complex potentials that reproduce large QM datasets. In this discussion I will discuss two different ways that we are pursuing this goal. The first approach uses the ReaxFF family of potentials that reproduces known chemical reaction pathways in small clusters of atoms while still allowing molecular dynamics simulations of millions of atoms undergoing chemical reaction. The second approach, SNAP, is a very general machine-learning approach for automated generation of interatomic potentials from large QM datasets. I. Initiation in energetic materials is fundamentally dependent on the interaction between a host of complex chemical and mechanical processes, occurring on scales ranging from intramolecular vibrations through molecular crystal plasticity up to hydrodynamic phenomena at the mesoscale. A variety of methods [e.g. quantum electronic structure methods (QM), nonreactive classical molecular dynamics (MD), mesoscopic continuum mechanics] exist to study processes occurring on each of these scales in isolation, but cannot describe how these processes interact with each other. In contrast, the ReaxFF reactive force field, implemented in the LAMMPS parallel MD code, allows us to routinely perform multimillion-atom reactive MD simulations of shock-induced initiation in a variety of energetic materials. II. The growing availability of large QM data sets has made it possible to use automated machine-learning approaches for interatomic potential development. Bartok et al. demonstrated that the bispectrum of the local neighbor density provides good regression surrogates for QM models using Gaussian process regression. We adopted a similar bispectrum representation within a linear regression scheme that we called SNAP. We have produced potentials for tantalum and indium phosphide. Results will be presented demonstrating the accuracy of these potentials relative to the training data, as well as their ability to accurately predict material properties not explicitly included in the training data. Comparing to recent QM calculations of screw dislocation cores in BCC tantalum, we observe that the SNAP potential gives the correct core structure and the correct energy barrier for screw dislocation motion, unlike existing EAM and ADP potentials.