Towards robust error control for high-throughput materials simulations

The discovery and development of new materials nowadays heavily depends on systematic high-throughput simulations. The resulting data is either directly employed to filter a large design space for promising compounds or to train sophisticated surrogates. In this setting questions related to the robustness and accuracy of simulation protocols as well as the reproducibility of obtained simulation data are more pressing then ever.

In this talk I will provide an overview of the research in the Mathematics for Materials Modelling group to improve efficiency and reliability of density-functional theory (DFT) simulations --- the most common physical model employed in the field. This includes algorithms that automatically adapt to the physics of the modelled system or multi-tasking surrogates, which can be trained from data of heterogeneous quality.

Originating from the interdisciplinary nature of our research we also develop software, that fosters cross-disciplinary collaboration. Our workhorse is the density-functional toolkit (DFTK, https://dftk.org), a Julia-based software for DFT simulations. With only 7500 lines DFTK is accessible to mathematical research as sketched above. At the same time the code is efficient and well-integrated with high-throughput libraries such as Aiida. New methods can therefore be directly employed in the usual setting for tackling research questions in materials science.