Speaker
Description
Density functional theory (DFT) is routinely used to study matter at ambient and extreme conditions. Studying matter at extreme conditions is computationally expensive, since the degrees of freedom and consequently the configurational space increases concurrently with increasing temperature and pressure. Machine learning-based interatomic potentials (ML-IAP) provide a feasible solution to extend the length and timescales of simulations, paving way for exploring states of matter not possible using ab-initio methods. The majority of existing descriptors required to construct ML-IAP neglect the spin degrees of freedom due to additional complexity. Here, we present our preliminary ideas/workflows to construct "spin-aware" ML-IAP using the SNAP descriptors [1] for heavy elements using training data sets obtained from density functional theory-molecular dynamics (DFT-MD) and machine learning techniques [2,3]. We further discuss the transferibility of ML-IAP from describing spin-phases in solids at ambient and sub-extreme conditions to liquids at extreme conditions.
[1] Thompson, Aidan P., et al. "Spectral neighbor analysis method for automated generation of quantum-accurate interatomic potentials." Journal of Computational Physics, 285, 316 (2015).
[2] Nikolov, Svetoslov, et al. "Data-driven magneto-elastic predictions with scalable classical spin-lattice dynamics." NPJ Computational Materials, 7, 153 (2021).
[3] Nikolov, Svetoslov, et al. "Dissociating the phononic, magnetic and electronic contributions to thermal conductivity: a computational study in alpha-iron" Journal of Materials Science (in press) (2022).
