| About this Abstract |
| Meeting |
2021 TMS Annual Meeting & Exhibition
|
| Symposium
|
Materials and Chemistry for Molten Salt Systems
|
| Presentation Title |
Fast and Accurate High-dimensional Neural Network Interatomic Potentials for Lithium-based Fluoride Salts |
| Author(s) |
Stephen Lam, Qing-Jie Li, Ronald Ballinger, Charles Forsberg, Ju Li |
| On-Site Speaker (Planned) |
Stephen Lam |
| Abstract Scope |
Experimental data are difficult to collect for molten salts due to the complexities and cost in handling liquids at high temperatures. Computational modeling can elucidate physical mechanisms driving behavior and can thus provide predictive capacity and help guide experimental trials. In this work, density functional theory (DFT) was used to train neural networks potentials, which were used to run molecular dynamics simulations for common constituent LiF and prototypical salt Flibe (66.6%LiF-33.3BeF2). Neural network potentials accelerated computation by several orders of magnitude and allowed exploration of longer time scales and larger system sizes beyond hundreds of atoms. It was shown that neural network potentials can accurately reproduce multi-component liquid salt system energies with a mean average error of < 2 meV/atom, and atomic forces with an error of < 0.06 eV/Å. This enabled fast and accurate prediction of chemistry, transport properties, and local structure. |
| Proceedings Inclusion? |
Planned: |
| Keywords |
Modeling and Simulation, Nuclear Materials, Machine Learning |