Neural Network Interatomic Potentials
Abstract
Modern large-scale atomistic simulations of materials are enabled by classical inter- atomic potentials. Such potentials directly relate atomic coordinates to the energy and forces acting on atoms without performing first-principles calculations. Traditional potentials are based on physical models of interatomic bonding with free parameters fit to a limited set of experimental and first- principles data. In recent years, a new approach has emerged in which the potentials are developed by numerical interpolation between the points of a large first-principles database by applying machine-learning methods. In particular, arti#cial neural networks have been utilized to construct several machine-learning potentials. Being purely mathematical, such potentials enable extremely accurate predictions of energy and forces but suffer from poor transferability outside the training dataset. This project explores a different direction in which neural networks will be combined with physics-based models of interatomic bonding. By contrast to the mathematical potentials, the proposed physical neural-network (PNN) potentials are expected to demonstrate muchbetter transferability while drastically improving the accuracy of the traditional potentials. PNN potentials have been constructed for Si and Al and their advantages over the mathematical neural network potentials have been demonstrated. At the next stage of the project, PNN copper, tungsten and bismuth potentials will be created. The work on single-component potentials will prepare the ground for extension of the PNN formalism to binary and multi-component systems in the future.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jul 26, 2018
- Source ID
- N000141812612
Entities
People
- Y. Mishin
Organizations
- George Mason University
- Office of Naval Research
- United States Navy