Abstract:

　　Important properties of materials, such as crystal structure, activity and selectivity of a catalyst, or a thermoelectric’s figure of merit, are in general difficult to predict, in particular from first principles. The problem lies in the extreme complexity of the relation between the atomic composition of a material and its functional properties. We demonstrate how to bridge this complexity with artificial intelligence (AI). Although first-principles calculations can have the required accuracy to predict crystal structure, the computational cost is too high for screening several different structures for many materials. We show how a compressed-sensing approach LASSO[1] and its further development SISSO[2] allow for finding easily computable descriptive parameters (descriptors), which can be used to quickly screen relative stability of different crystal structures across the chemical space. We demonstrate that the found physical descriptors allow us to predict crystal structure of materials with chemical compositions that were not included in the training of AI. Employing SISSO, we also find descriptors for hydrogen molecule activation on single-atom alloy catalysts, and screen for the best catalyst for hydrogenation reactions among thousands of candidates.[3]

References:

[1] Ghiringhelli, L. M. et al., Phys. Rev. Lett., 114, 105503 (2015).

[2] Ouyang, R. et al., Phys. Rev. Mater., 2, 083802 (2018).

[3] Han, Z. et al., Nat. Comm. 12, 1833 (2021).

Biosketch:

　　Sergey V. Levchenko obtained M.Sc. from the Moscow Institute of Physics and Technology, and Ph.D. from Univesity of Southern California, LA, USA, in 2005. After a postdoc period at the University of Pennsylvania, PA, USA, he was a group leader at the Fritz Haber Institute of Max Planck Society in Berlin, Germany, and since 2018 he is a professor at Skoltech, Moscow, Russia. Sergey Levchenko is an expert in materials modelling using first-principles methods and artificial intelligence.