3DSC - a dataset of superconductors including crystal structures

Timo Sommer, Roland Willa, Jörg Schmalian, Pascal FriederichScientific 2023 · DOI

Abstract

Data-driven methods, in particular machine learning, can help to speed up the discovery of new materials by finding hidden patterns in existing data and using them to identify promising candidate materials. In the case of superconductors, the use of data science tools is to date slowed down by a lack of accessible data. In this work, we present a new and publicly available superconductivity dataset ('3DSC'), featuring the critical temperature $T_c$ of superconducting materials additionally to tested non-superconductors. In contrast to existing databases such as the SuperCon database which contains information on the chemical composition, the 3DSC is augmented by approximate three-dimensional crystal structures. We perform a statistical analysis and machine learning experiments to show that access to this structural information improves the prediction of the critical temperature $T_c$ of materials. Furthermore, we provide ideas and directions for further research to improve the 3DSC. We are confident that this database will be useful in applying state-of-the-art machine learning methods to eventually find new superconductors.

Introduction

Predicting the critical temperature ($T_c$) at which a material becomes a superconductor is a central challenge in materials science. While the SuperCon database provides a massive repository of superconducting materials, it primarily focuses on chemical formulas. However, $T_c$ is highly sensitive to the exact 3D arrangement of atoms. Polymorphs—materials with identical formulas but different crystal structures—can exhibit vastly different superconducting properties. Without 3D structural data, machine learning (ML) models struggle to account for the microscopic physics, such as phonon spectra and electron-lattice coupling, that drive superconductivity.

The 3DSC Pipeline: From Formulas to Structures

To address this data gap, the researchers developed a pipeline to "structure-augment" the SuperCon database. This was achieved by matching SuperCon entries with crystal structures from the Materials Project (MP) and the Inorganic Crystal Structure Database (ICSD). The core of this process is a matching and adaptation algorithm designed to handle cases where exact structural matches are unavailable.

Schematic of the matching and adaptation algorithm used to generate the 3DSC dataset.

The pipeline utilizes formula normalization to match materials that differ only by a scaling factor. When an exact match is missing, the algorithm employs "artificial doping," searching for a proxy structure with a similar formula and replacing atoms at equivalent crystal sites to match the target composition.

Artificial Doping and Physical Constraints

The artificial doping approach significantly expands the dataset's reach. By allowing for proxy structures, the researchers increased the number of usable data points by 773% for the ICSD and 660% for the Materials Project. To maintain physical plausibility, the algorithm adheres to strict constraints: replacements must be unambiguous, involve no more than two elements per site, and preserve the original number of atomic sites.

Example of a successful artificial doping match where Rh is substituted by Ir at equivalent sites.

Machine Learning Performance and Insights

The team validated the 3DSC dataset using Gradient Boosting (XGBoost) models. Key findings include:

  • Structural Regularization: Incorporating structural features (DSOAP) increased training error but decreased test error. This suggests that structural data acts as a regularizer, forcing models to learn general physical relationships rather than simply memorizing chemical formulas.
  • Improved Accuracy: For the $3DSC_{MP}$ dataset, adding structural features improved the Mean Squared Logarithmic Error (MSLE) from 0.776 to 0.748.
  • Data Efficiency: Even with only 36% of the original SuperCon data, the structure-aware model outperformed formula-only models trained on the entire database.
  • Class-Specific Benefits: High-$T_c$ superconductors like cuprates showed the most significant improvements, as their properties are deeply tied to specific oxygen-plane arrangements.

Limitations and Future Directions

While 3DSC provides a powerful new tool, the matched structures remain approximations. Future work will focus on incorporating electronic data (e.g., band gaps and Fermi energy), refining the ranking criteria for multiple structural matches, and integrating additional sources like the Crystallography Open Database (COD).

Summary updated

Abstract

Data-driven methods, in particular machine learning, can help to speed up the discovery of new materials by finding hidden patterns in existing data and using them to identify promising candidate materials. In the case of superconductors, the use of data science tools is to date slowed down by a lack of accessible data. In this work, we present a new and publicly available superconductivity dataset ('3DSC'), featuring the critical temperature T<sub>C</sub> of superconducting materials additionally to tested non-superconductors. In contrast to existing databases such as the SuperCon database which contains information on the chemical composition, the 3DSC is augmented by approximate three-dimensional crystal structures. We perform a statistical analysis and machine learning experiments to show that access to this structural information improves the prediction of the critical temperature T<sub>C</sub> of materials. Furthermore, we provide ideas and directions for further research to improve the 3DSC. We are confident that this database will be useful in applying state-of-the-art machine learning methods to eventually find new superconductors.
BibTeX
@inproceedings{sommer2023dsc,
  author = {Timo Sommer and Roland Willa and Jörg Schmalian and Pascal Friederich},
  title = {{3DSC - a dataset of superconductors including crystal structures}},
  booktitle = {Scientific},
  year = {2023},
  doi = {10.1038/s41597-023-02721-y},
}
Figure