This title appears in the Scientific Report :
2023
Please use the identifier:
http://dx.doi.org/10.34734/FZJ-2023-05856 in citations.
Benchmark study of symmetry-adapted ML-DFT models for magnetically doped topological insulators
Benchmark study of symmetry-adapted ML-DFT models for magnetically doped topological insulators
We present a benchmark study of surrogate models for impurities embedded into crystalline solids. Using the Korringa-Kohn-Rostoker Green Function method [1], we have built databases of several thousand calculations of single impurities (monomers) embedded into different elemental crystals, as well a...
Saved in:
Personal Name(s): | Wasmer, Johannes (Corresponding author) |
---|---|
Mozumder, Rubel / Rüssmann, Philipp / Assent, Ira / Blügel, Stefan | |
Contributing Institute: |
Quanten-Theorie der Materialien; IAS-1 Quanten-Theorie der Materialien; PGI-1 |
Imprint: |
2022
|
DOI: |
10.34734/FZJ-2023-05856 |
Conference: | Psi-k 2022 Conference, Lausanne (Switzerland), 2022-08-22 - 2022-08-25 |
Document Type: |
Poster |
Research Program: |
Joint Virtual Laboratort for AI, Data Analytics and Scalable Simulation Helmholtz School for Data Science in Life, Earth and Energy (HDS LEE) Topological Matter |
Link: |
Get full text OpenAccess |
Publikationsportal JuSER |
We present a benchmark study of surrogate models for impurities embedded into crystalline solids. Using the Korringa-Kohn-Rostoker Green Function method [1], we have built databases of several thousand calculations of single impurities (monomers) embedded into different elemental crystals, as well as of the topological insulator Bi2Te3, magnetically co-doped with transition metal impurities (dimers). We predict the converged monomer impurity electron potential and the isotropic exchange interaction of the impurity dimer in the classical Heisenberg model. From these surrogates, we intend to build transferable models for larger systems in the future, which will accelerate the convergence of our DFT codes. The study compares various recent E(3)-equivariant models such as ACE and NequIP [2] in terms of performance and reproducible end-to-end workflows.[1] P. Rüßmann et al., npj Comput Mater 7, 13 (2021)[2] I. Batatia et al., arXiv:2205.06643 (2022) |