Towards multiscale ab-initio simulations: size transferability of density functional theory surrogates

Density Functional Theory (DFT) is one of the most important computational tools for materials science, as it combines high accuracy with general computational feasibility. However, applications important to scientific progress can pose problems to even the most advanced and efficient DFT codes due to size and/or complexity of the underlying simulations. Namely the modeling of materials across multiple length and time scales at ambient or extreme conditions, necessary for the understanding of important physical phenomena such as radiation damage in fusion reactor walls, evade traditional ab-initio treatment.
DFT surrogate models are a useful tool in achieving this goal by reproducing DFT results at drastically reduced computational cost due to using machine learning methods. In order to successfully model on multiple length and time scales, these models have to be transferable with respect to their size. Here, we present results of such an investigation, by showing how models trained on small numbers of atoms (e.g., 128) can be used to accurately calculate energies of much larger simulation cells (e.g., 1024 atoms). The models are based upon the Materials Learning Algorithms (MALA) package and the LDOS-based machine-learning workflow implemented therein.