Machine learning-assisted screening of intrinsic rattling compounds with large atomic displacement

Abstract

Thermal conductivity is a key thermophysical property governing the heat transport in materials. Specifically, some applications such as thermoelectrics and thermal coatings need ultralow thermal conductivity. In this work, we established a correlation between machine learning models and the mean square displacement with high efficiency and accuracy considering that large atomic displacements can be regarded as reasonable criteria for ultralow thermal conductivity. The results show that the prediction performance of traditional machine learning models, such as random forest, which are based solely on composition-weighted elemental properties, is comparable to that of advanced graph neural network models. Deep insight into the underlying physical and chemical properties reveals that atomic features related to the volume and bonding strength demonstrate a close correlation with the mean square displacement. By projecting onto the space of significant atomic features, the constituent elements and structure prototypes that have the potential for substantial atomic displacement are identified. In particular, halide double perovskites are reported to be promising structures exhibiting large atomic displacement. To verify the prediction results, the mean square displacements of 20 halide double perovskites are further validated by first-principles calculations, and intrinsic rattling vibrations are also recognized in this structure prototype. This work proposes a viable method for the rapid screening of materials with considerable atomic displacement based on simple elemental and structural properties, thereby facilitating the discovery of potential candidates for ultralow thermal conductivity.