Discovering Better Catalysts from Artificial Intelligence Analysis of Ab Initio Data: Application to the CO2 Conversion Problem
Aliaksei Mazheika1, Yanggang Wang1,2, Luca Ghiringhelli1, Sergey V. Levchenko3,1,4, Matthias Scheffler1
1Fritz-Haber-Institute of the Max-Planck-Society, Berlin, Germany
2University of Shenzhen, Shenzhen, China
3Skolkovo Innovation Center, Moscow, Russia
4NUST MISIS, Moscow, Russia
Email: mazheika@fhi-berlin.mpg.de
Using artificial intelligence (AI) trained on ab initio data, we develop a strategy for the rational
design of catalytic materials for converting CO2 to fuels and other useful chemicals. Specifically,
we employ the subgroup discovery [1] and sure independence screening and sparsifying operator
(SISSO) [2]. For oxide surfaces the results reveal that an electron transfer to the π*-antibonding
orbital of the adsorbed molecule and the associated bending of the initially linear O-C-O,
previously proposed as indicator of activation [3], are insufficient to account for the good
catalytic performance. Instead, our AI model identifies the common feature of a group of
experimentally studied oxide catalysts in the binding of one molecular O atom to a surface
cation, which results in a strong elongation and therefore weakening of the molecular C-O bond.
This suggests using the C-O bond elongation as an indicator of CO2 activation. Based on these
findings, we propose a set of new promising catalysts for CO2 conversion, and a recipe to find
more.
- M. Boley et al. Data Min. Knowl. Disc. 31, 1391 (2017).
- R. Ouyang et al. Phys. Rev. M. 2, 083802 (2018).
- H.-J. Freund and M. W. Roberts, Surf. Sci. Rep. 25, 225 (1996).