Speaker
Description
We present Prokyon [1], a free software library that enables on-the-fly machine learning (ML) of interatomic potentials for atomistic simulations that can be efficiently coupled to existing electronic structure packages. Apart from the energy and forces, our model also predicts the uncertainty in these quantities, which is a major advantage of the Gaussian process (GP) approach used here when compared to, e. g., neural network models. By doing so, we can judge if a certain simulation step can be reliably predicted via ML. The ML model is initialized from scratch and trained iteratively on-the-fly in parallel with an ab initio molecular dynamics (AIMD) run, so that more and more steps can be reliably predicted as the simulation proceeds, leading to an increasing speedup of the simulation. Prokyon is designed as a modular framework for incorporating different ML approaches. The library currently includes an optimized implementation of the FLARE [2] model based on atomic cluster expansion (ACE) [3] descriptors and sparse Gaussian process regression (GPR). Initial testing shows that we can reach an accuracy which is at least on par with other existing on-the-fly ML approaches such as the one implemented in VASP [4]. Prokyon can be integrated into existing AIMD program packages, so that all the methods available in these packages can then be coupled to our ML proto-col, opening the door to time scales which were previously out of scope for AIMD simulations. As a first step, Prokyon will be integrated into the CP2k [5] and ORCA [6] packages.
[1] https://prokyon-lib.org/
[2] J. Vandermause, S. B. Torrisi, S. Batzner, Y. Xie, L. Sun, A. M. Kolpak, B. Kozinsky,
npj Comput. Mater. 2020, 6, 20.
[3] R. Drautz, Phys. Rev. B 2019, 99, 014104.
[4] R. Jinnouchi, F. Karsai, G. Kresse, Phys. Rev. B 2019, 100, 014105.
[5] https://www.cp2k.org/
[6] https://orcaforum.kofo.mpg.de/