Atomic Distance Kernel for Material Property Prediction
A comprehensive search of various candidate materials is an important step in discovering novel materials with desirable physical properties. However, the search space is quite vast, so that it is not practical to perform exhaustive experiments to check all the candidates. Even if the chemical composition is the same, the properties of materials may differ significantly depending on the crystal structure, and therefore, the number of possible combinations increases considerably. Recently, machine learning methods have been successfully applied to material search to estimate prediction models using existing databases and predict the physical properties of unknown substances. In this research, we propose a novel kernel function between compounds, which directly uses crystal structure information for the prediction of physical properties of inorganic crystalline compounds based on the crystal structures. We conduct evaluation experiments and show that the structure information improves the prediction accuracy.
KeywordsMachine learning Kernel Material informatics
This work was supported by MEXT Grant-in-Aid for Scientific Research on Innovative Areas, Exploration of Nanostructure-property Relationships for Materials Innovation.
- 3.Nakayama, K.: Compound descriptors for machine learning based material design. Master’s thesis, Department of Materials Science and Engineering, Graduate School of Engineering, Kyoto University (2015)Google Scholar
- 4.Rupp, M., Tkatchenko, A., Müller, K.R., Lilienfeld, V., Von Lilienfeld, O.A.: Fast and accurate modeling of molecular atomization energies with machine learning. Phys. Rev. Lett. 108(5), 58301 (2012). http://eprints.pascal-network.org/archive/00009418/ CrossRefGoogle Scholar