Comparison of structural representations for machine learning-accelerated ab initio calculations

Poster presented at workshop "84. Annual Meeting of DPG and DPG-Tagung (DPG Meeting) of the Condensed Matter Section" (SKM21), 2021-09-27 to 2021-10-01.

• Event website
• Abstract on event website

Abstract.

Quantum mechanical calculations based on density functional theory (DFT) are the workhorse in today’s computational materials design. Here we explore the possibility to accelerate the DFT calculations with potentials generated from a surrogate machine learning model. Finding a better starting potential could drastically reduce the number of required self-consistency steps during the convergence of DFT calculations. The juKKR code (jukkr.fz-juelich.de) allows high-throughput ab initio impurity embedding calculations which we use to generate a training dataset of 10’000 impurities from most elements of the periodic table embedded into elemental crystals with the help of the workflow engine AiiDA. The choice of a structural representation of the atomic environment which a machine learning model can understand has been identified as a crucial step. We compare a variety of such representations as training input for our surrogate model. Finally, we benchmark results for the converged impurity potential from DFT calculations against the output of the trained surrogate model.

We acknowledge support by the Joint Lab Virtual Materials Design, by the DFG under Germany’s Excellence Strategy Cluster of Excellence ML4Q, by the AIDAS2 virtual lab, and thank for computing time provided on the JARA Partition part of the supercomputer CLAIX at RWTH Aachen University.