Superconductivity in antiperovskites

Perovskite materials (XYZ3), a widely studied family of ternary compounds, find applications across various technological domains, including photovoltaics, piezoelectricity, magnetism, and superconductivity. The discovery of superconductivity at 8 K in the non-oxide perovskite MgCNi3 in 2001 sparked significant interest. Initially attributed to unconventional mechanisms due to the high Ni content, MgCNi3 is now recognized as an s-wave superconductor with electron-phonon interaction mediated pairing. Subsequent research identified several superconducting carbides, borides, nitrides, and oxide antiperovskites (inverted perovskites, with the nonmetal in the Y position). Experimental maximum Tc reached 10 K (InBLa3 and InOLa3), while theoretical predictions suggested higher values for certain compounds. This work focuses on conventional superconductivity in antiperovskites, aiming to understand the behavior of the entire compound family through a large-scale study, combining density-functional perturbation theory for electron-phonon properties with interpretable machine-learning models for prediction and interpretation of data.

The literature review section extensively cites previous research on perovskite and antiperovskite materials, highlighting the known superconducting compounds and theoretical predictions. It discusses the differences between standard perovskites and antiperovskites, focusing on the position of the nonmetal atom in the crystal structure. The review emphasizes the rarity of large-scale studies on entire material families and the novelty of combining standard computational methods with machine learning techniques to predict and interpret superconducting properties.

The researchers selected all inverted perovskites with H, B, C, N, O, and P from a systematic study, filtering out semiconductors and magnetic materials. They focused on compounds within 50 meV/atom of the convex hull to ensure thermodynamic feasibility. Density-functional perturbation theory (using QUANTUM ESPRESSO with pseudopotentials from PSEUDODOJO) was employed to calculate electron-phonon coupling strength (λ), average phonon frequency (ωlog), and Tc. Two machine learning algorithms were utilized: SISSO (combining symbolic regression and compressed sensing for interpretable models) and MAPLE (providing accurate predictions and interpretability through local linear models). Input features included structure volume, atomic charges, density-of-states at the Fermi level, and atomic properties (row, column in periodic table, electronegativity, atomic weight, covalent radius). Cross-validation was used to assess the models' performance, dividing the dataset into training (80%) and validation (20%) sets. The models were trained to predict λ and ωlog, and the results were then used to predict Tc for a broader range of antiperovskites (within 400 meV/atom of the convex hull). The methodology also incorporated analysis of the effects of pressure on Tc for selected materials.

The high-throughput calculations revealed that most of the 397 studied systems exhibited weak electron-phonon coupling and Tc < 1 K. However, 16 compounds displayed Tc > 5 K, including antiperovskites with Y = H, N, C, and O. The five materials with the highest predicted Tc are listed in the paper, along with a complete list in supplementary information. The machine learning models provided interpretable formulas linking λ and ωlog to material properties, showing λ is proportional to the density-of-states at the Fermi level and inversely proportional to the mass of the Y atom, while ωlog is mainly determined by the Z atom. Expanding the search to materials farther from the convex hull (up to 400 meV/atom), using the machine learning models, identified an additional 55 materials with Tc > 5 K, reaching a maximum of 17.8 K for PtHBe3. Analysis of specific materials (PtHBe3, ScCRh3, MoNMn3, AsHTi3) revealed details of their electronic and phononic properties and their behavior under pressure. The findings indicated that the phonon modes responsible for Cooper pair binding are related to vibrations of the Z atoms (forming the octahedra), these modes being rather soft, enhancing λ but lowering ωlog. A balance is needed to obtain high Tc, avoiding instability.

The study's findings address the research question by providing a comprehensive understanding of superconductivity in antiperovskites, highlighting the importance of the density-of-states at the Fermi level, and the interplay between the atomic masses and properties of the constituent elements in determining Tc. The significance of the results lies in the successful combination of traditional density-functional theory with machine-learning models to predict and interpret superconductivity in a large family of materials. The approach has potential to accelerate material discovery for applications in superconductivity. The results emphasize the importance of considering thermodynamic stability and the potential for synthesizing metastable materials.

This research demonstrates the power of combining density functional theory with interpretable machine learning to study superconductivity in antiperovskites. The identification of numerous promising candidates with Tc > 5K and a maximum of 17.8K for PtHBe3 highlights the potential of this approach for materials discovery. Future work could focus on experimental synthesis and characterization of the predicted materials, exploring the effects of defects and anharmonicity, and extending the methodology to other material families and properties.

The study relies on theoretical calculations and predictions; experimental verification of the predicted superconducting properties is necessary. The models may not capture all factors influencing superconductivity, such as anharmonicity and defects, limiting the accuracy of predictions for materials far from thermodynamic stability. The accuracy of the predictions is dependent on the quality of the input data and the limitations of density-functional theory itself.