This title appears in the Scientific Report :
2023
Please use the identifier:
http://dx.doi.org/10.34734/FZJ-2023-05853 in citations.
Comparison of structural representations for machine learning-enhanced DFT of impurity embeddings
Comparison of structural representations for machine learning-enhanced DFT of impurity embeddings
The acceleration or even replacement of ab initio methods for atomistic systems with surrogate models based on machine learning has gained traction in recent years [1]. This development stands on two pillars: The first one is the fast growth of materials databases, thanks in part to high-throughput...
Saved in:
Personal Name(s): | Wasmer, Johannes (Corresponding author) |
---|---|
Rüssmann, Philipp / Blügel, Stefan | |
Contributing Institute: |
Quanten-Theorie der Materialien; PGI-1 Quanten-Theorie der Materialien; IAS-1 |
Imprint: |
2022
|
DOI: |
10.34734/FZJ-2023-05853 |
Conference: | Virtual Materials Design 2021, online (Germany), 2021-07-20 - 2021-07-21 |
Document Type: |
Poster |
Research Program: |
Joint Virtual Laboratort for AI, Data Analytics and Scalable Simulation EXC 2004: Matter and Light for Quantum Computing (ML4Q) EXC 2004: Materie und Licht für Quanteninformation (ML4Q) Topological Matter |
Link: |
Get full text Restricted OpenAccess |
Publikationsportal JuSER |
The acceleration or even replacement of ab initio methods for atomistic systems with surrogate models based on machine learning has gained traction in recent years [1]. This development stands on two pillars: The first one is the fast growth of materials databases, thanks in part to high-throughput calculation (HTC) infrastructures such as AiiDA [2]. The second one is advances in method development in atomistic machine learning, where finding the best representation of an atomic system as input for model training has been identified as a crucial step to success. Structural representations rely, like the Schrödinger equation, only on the atom positions and their chemical identity within a system [3], and are thus most suitable for this task.Here we investigate the possibility to accelerate the density functional theory (DFT) code juKKR [4] with machine learning starting potentials. This code has been used for instance to perform HTC on impurity embeddings into topological insulators [5]. We use a combinatorial approach to generate 7000 impurity embeddings from most elements of the periodic table into elemental crystals with the help of AiiDA. We generate their fingerprints using structural descriptors implemented in the DScribe package [6], such as smooth overlap of atomic positions. To benchmark their representational power for these embeddings, we present the results of a simple classification experiment.We acknowledge support by the Joint Lab Virtual Materials Design (JL-VMD) and thank for computing time granted by the JARA Vergabegremium and provided on the JARA Partition part of the supercomputer CLAIX at RWTH Aachen University. This work was funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) under Germany's Excellence Strategy – Cluster of Excellence Matter and Light for Quantum Computing (ML4Q) EXC 2004/1 – 390534769, and by AIDAS2 – AI, Data Analytics and Scalable Simulation – a virtual lab between CEA, France and FZJ, Germany. |