Yuqing He^1,2, Pierre-Paul De Breuck¹, Hongming Weng², Matteo Giantomassi¹, Gian-Marco Rignanese^1*

1 UCLouvain, Institut de la Matiere Condensée et des Nanosciences (IMCN), Chemin des Étoiles 8, Louvain-la-Neuve 1348, Belgium

2 Beijing National Laboratory for Condensed Matter Physics and Institute of Physics,Chinese Academy of Sciences, Beijing, China

* Corresponding authors emails: gian-marco.rignanese@uclouvain.be

DOI10.24435/materialscloud:zk-gc [version v2]

Publication date: Feb 26, 2025

How to cite this record

Yuqing He, Pierre-Paul De Breuck, Hongming Weng, Matteo Giantomassi, Gian-Marco Rignanese, Machine learning on multiple topological materials datasets, Materials Cloud Archive 2025.32 (2025), https://doi.org/10.24435/materialscloud:zk-gc

Description

A dataset of 35,608 materials with their topological properties is constructed by combining the density functional theory (DFT) results of Materiae and the Topological Materials Database. Thanks to this, machine-learning approaches are developed to categorize materials into five distinct topological types, with the XGBoost model achieving an impressive 85.2% classification accuracy. By conducting generalization tests on different sub-datasets, differences are identified between the original datasets in terms of topological types, chemical elements, unknown magnetic compounds, and feature space coverage. Their impact on model performance is analyzed. Turning to the simpler binary classification between trivial insulators and nontrivial topological materials, three different approaches are also tested. Key characteristics influencing material topology are identified, with the maximum packing efficiency and the fraction of p valence electrons being highlighted as critical features.

Materials Cloud sections using this data

Files

File name	Size	Description
topoclass.ipynb MD5md5:e97367839d6f923849f6ab81ff7830f8	48.8 KiB	Python notebook to analyze the data
topoclass.json.gz MD5md5:57066e24446f9f128c63a8e1698880de	515.6 MiB	JSON gzipped file with all the data
topoclass.chemiscope.json.gz MD5md5:d9c6901e2781b55382233974c5360bfc Visualize on Chemiscope	19.8 MiB	Chemiscope visualization of the dataset, that can be used interactively on Materials Cloud
optimade.jsonl MD5md5:2859bad2900c0ad1057e8b3d357f452e	101.8 MiB	OPTIMADE file with all the data
optimade.yaml MD5md5:2b92ac74ef6a143d4ca43807b007da2a Go to the OPTIMADE API	237 Bytes	OPTIMADE yaml configuration file

License

Files and data are licensed under the terms of the following license: Creative Commons Attribution 4.0 International.
Metadata, except for email addresses, are licensed under the Creative Commons Attribution Share-Alike 4.0 International license.

External references

Keywords

topological materials machine learning database