Towards multiscale simulations for matter under extreme conditions: Building surrogate models with machine learning

The accurate numerical treatment of matter under extreme conditions is crucial for the understanding of important physical phenomena such as radiation damage in fusion reactor walls, or planetary interiors. Yet, such simulations are unfeasible with state-of-art methods, e.g., density functional theory (DFT) if performed at large length and time scales, due to unfavorable scaling behavior. One possible route to mitigate these scaling issues are machine-learning based surrogate models; DFT data is used to calculate models that allow access to the same quantities of interest a DFT simulation would, at drastically reduced computational cost. CASUS (in cooperation with SNL and ORNL) develops a framework called "Materials Learning Algorithms" (MALA), drawing on which DFT surrogate models can easily be created and applied. Here we present an overview of MALA and recent results, such as size transferability and automated model construction.