A Flexible and Scaleable Scheme for Combining Formation Energies Computed with Different Density Functionals

Formation energy, decomposition energy, and other energy-derived properties are central to computational materials screening efforts, which often entail performing density functional theory (DFT) calculations for hundreds or thousands of distinct materials. Historically, the vast majority of such calculations have employed popular GGA functionals (notably, PBE) due to their efficient compromise between accuracy and computational cost. However, high-throughput calculations at higher levels of theory (e.g., metaGGA) are becoming feasible and even routine thanks to continued advances in theory and computing power. Nevertheless, recomputing all of the hundreds of thousands of GGA calculations that currently populate large materials datbases such as the Materials Project or NOMAD at a higher level of theory would be an unwise and arguably unnecessary use of resources. Instead, we propose targeting higher level of theory calculations to materials in which they are most likely to improve accuracy compared to GGA.<br/><br/>To make this possible, we have developed a scheme by which DFT energies computed at different levels of theory may be combined onto a single phase diagram to generate improved estimates of formation energy. In this presentation, we will demonstrate the scheme using GGA and metaGGA calculations generated via high-throughput computational workflows adopted by the Materials Project; however the scheme may be used to robustly combine energies from any two DFT functionals or levels of theory. We will show that through careful consideration of thermodynamic reference states and convex energy hull construction, our flexible scheme enables systematic improvements in formation energy predictions using as few as 2 metaGGA calculations, and scales with the number of metaGGA calculations all the way to the limit in which an entire chemical system has been recomputed. In addition to solid formation energies, mixed energies may be used to improve aqueous phase stability prediction via computational Pourbaix diagrams, as we will show.<br/><br/>By providing a way to leverage a relatively small number of higher-level calculations to improve formation energy predictions in large chemical systems, our robust, scaleable, and flexible mixing scheme lays the foundation for a step change in the accuracy of high-throughput materials screening efforts.