Accelerating ab initio molecular dynamics simulations with on-the-fly machine learned force fields
Author(s)
Kumar, Shashikant
Advisor(s)
Phanish Suryanarayana
Editor(s)
Collections
Supplementary to:
Permanent Link
Abstract
A comprehensive framework is presented for on-the-fly machine learned force field (MLFF) to accelerate the ab initio molecular dynamics (MD) simulations. The framework utilizes an MLFF scheme based on the kernel method and Bayesian linear regression, leveraging Kohn-Sham density functional theory (DFT) training data. The SOAP descriptor framework is used to fingerprint the local atomic environments. We study the efficiency and accuracy of the MLFF scheme through various examples. We apply the framework of on-the-fly MLFF to study the transport properties of CH in warm dense matter (WDM) conditions. Specifically, we calculate the diffusion coefficients and shear viscosity of CH in these conditions. Notably, a comparison with previous results showcases agreement within the standard deviation for carbon and CH, underscoring the robustness of our approach. In terms of performance, we achieve a three-order of magnitude speedup over the ab initio MD simulation, obtaining a high accuracy in the diffusion coefficients and viscosity.
We Also introduce a delta-ML framework to enhance the accuracy of orbital-free DFT (OF-DFT) simulation. Specifically, we learn the differences in the energy, and forces between Kohn-Sham DFT (KS-DFT) and OF-DFT using the on-the-fly framework described above. We find that the formalism not only improves the accuracy of Thomas–Fermi–von Weizsacker orbital-free energies and forces by more than two orders of magnitude but is also more accurate than MLFFs based solely on KS-DFT while being more efficient and less sensitive to model parameters. We apply the framework to study the structure of molten Al0.88Si0.12, the results suggesting no aggregation of Si atoms, in agreement with a previous Kohn–Sham study performed at an order of magnitude smaller length and time scales. We also utilize this delta-ML framework to study the effect of exchange correlations on equation state for materials in WDM conditions.
Finally, we develop a cyclic and helical symmetry-informed formulation of MLFF. This formulation builds upon the previously implemented cyclic and helical symmetry-adapted framework of Kohn-Sham DFT, extending it to enable phonon calculations. The adoption of symmetry allows the on-the-fly MLFF framework to efficiently study nanostructures like nanotubes. We apply this formulation to investigate the vibrational properties of carbon nanotubes, achieving high accuracy in phonon predictions.
Sponsor
Date
2024-12-02
Extent
Resource Type
Text
Resource Subtype
Dissertation