Towards all-electron treatment in electronic structure machine learning

Electronic structure calculations with density functional theory (DFT) are today the most widely used technique for computational materials design. As materials design enters the artificial intelligence (AI) age, large databases of these calculations, such as the Materials Project, form the basis for creating foundation models. This can be seen in leaderboards such as Matbench Discovery 1. These AI foundation models mark a paradigm shift in computational materials design since they accelerate the time to solution by orders of magnitude. However, the current DFT databases and AI surrogate models are not accurate enough to design quantum materials. This problem can be solved by employing all-electron, full-potential, fully relativistic DFT, that can reach the required calculation precision. The JuDFT.de codes are among the few established implementations of such methods 2. They are the only all-electron DFT codes supporting high-throughput data generation that implement the FAIR data guidelines via their plugin integration with the AiiDA.net workflow engine.We present our progress on the open problem of surrogate AI models for all-electron DFT data for the JuDFT codes. We focus on the JuKKR code, and materials systems with defects, such as magnetically doped topological insulators 3. We predict the converged electron potential for accelerated DFT, and first-principles exchange interactions for accelerated spin dynamics.1 Riebesell, J. et al. Matbench Discovery – A framework to evaluate machine learning crystal stability predictions. Preprint at https://doi.org/10.48550/arXiv.2308.14920 (2024).2 Bosoni, E. et al. How to verify the precision of density-functional-theory implementations via reproducible and universal workflows. Nat Rev Phys 6, 45–58 (2024).3 Mozumder, R., Wasmer, J., Antognini Silva, D., Blügel, S. & Rüßmann, P. High-throughput magnetic co-doping and design of exchange interactions in topological insulators. Phys. Rev. Mater. 8, 104201 (2024).