How can we learn from high and medium throughput screening?
Matthias Arenz, University of Bern, Switzerland
This presentation will address our research efforts to study complex catalysts through experimental means, with the aid of medium-throughput screening coupled to machine learning (ML).
Initially, we employed Bayesian Optimization, a machine learning (ML) optimization scheme, to identify active electrocatalyst compositions for the oxidation of hydrogen in the presence of carbon monoxide. The results demonstrated that a rather small number of catalyst samples, less than 100 different compositions, is sufficient to identify highly active electrocatalyst compositions without the necessity of making any prior assumptions. Furthermore, the machine-learned model constructed to correlate particle composition with catalytic activity could be compared to density functional calculations conducted by our collaborators at Copenhagen University (Group of Jan Rossmeisl). This comparison facilitated an examination of trends derived from both experimental and computational studies. This concept was subsequently expanded to encompass the oxygen evolution reaction [2]. The key to this approach is that DFT calculations are based on very defined assumptions, whereas experimental data are used to construct a “black box” model without assumptions. By comparing both models, namely the composition-activity correlationships, it is possible to derive information about the reaction as well as the state of the active catalyst.
In the presentation, we discuss the general approach, as well as experimental considerations, and outline future directions.
References
[1] VA Mints, JK Pedersen, A Bagger, J Quinson, AS Anker, KMØ Jensen, J Rossmeisl, M Arenz, ACS Catalysis, 2023, 12 (18), 11263-11271;
[2] VA Mints, KL Svane, J Rossmeisl, M Arenz, ACS Catalysis, 2024, 14, 6936-6944