The development and understanding of new materials play a key role for major industrial and societal challenges of the 21st century. Over the past years more and more materials have been design fully in silico. At the core of this are automated workflows, which systematically perform calculations on thousands to million of compounds. Compared to the early years of materials modelling, which was centred around performing a small number of computations on hand-picked systems, such high-throughput screening approaches have much stronger requirements with respect to the robustness of employed algorithms: even a small percentage of erroneous or failing calculations equals a large number of cases, which need to be checked by a human (causing idle time) and redone thereafter (causing a noteworthy computational overhead).
Research in the MatMat group centres around overcoming these obstacles and accelerating computational materials discovery by providing efficient self-adapting simulations featuring a control of simulation error. Achieving these aspects not only contributes crucially to reducing the human time required to setup and verify calculations (the typical bottleneck in high-throughput workflows), but it also enables promising prospects such as active learning or adaptive numerical schemes, which automatically balance errors in order to obtain a simulation result along a path of least computational effort. Our main application are first-principle methods based on density-functional theory (DFT), but we furthermore work on a range of other problems, such as the modelling of quantum spin systems using tensor methods.
This work is inherently a multidisciplinary research effort such that we participate in a number of cross-disciplinary research initiatives and collaborations. A central component of our research efforts is therefore additionally to develop interdisciplinary software platforms which allow all involved communities to synergically join their forces. Building on top of these codes we explore emerging opportunities such as mixed-precision linear algebra or algorithmic differentiaton for materials modelling and investigate their perspectives to obtain more efficient simulations or to better integrate physical models with data-driven approaches.