Unsupervised Learning Discovery of Catalyst Materials for Carbon Dioxide to Methanol Synthesis

Novel and highly efficient heterogeneous catalysts are urgently needed to combat greenhouse gas emissions by converting CO₂ into useful products such as methanol. Despite advances in high-throughput computational and experimental materials screening, discovering effective catalysts remains challenging due to the demanding nature of quantum mechanical calculations, the slow, expensive nature of catalyst testing, and the vast materials space. Machine learning offers a powerful tool to predict materials activity [1], speeding up the theoretical predictions; however, the complexity of modern industrial catalysts with many active sites and facets needs to be addressed by a good descriptor.<br/><br/>Adsorption energies of reactants or intermediates have presented themselves as good proxies for the activity of basic metal catalysts [2]. However, a single adsorption energy can fail to describe the activity of more complex materials. Therefore, in this study, we introduce the adsorption energy distribution (AED) descriptor, extending the original concept by incorporating data on various facets and binding sites. We have compiled AEDs of relevant reaction intermediates, specifically for CO₂ to methanol thermal conversion reaction for approximately 200 materials utilizing a machine-learned interatomic potential from the Open Catalyst Project [3]. We have computed pairwise Wasserstein distances between the AEDs and performed hierarchical clustering to find groups of catalysts with comparable catalytic activities. The results show clustering and similar statistical markers for materials exhibiting high methanol yield in the experiment as well as suggesting new potential candidate materials. Notably, our approach identifies similarities between the well-known Cu-Zn-based thermal catalyst [4] and materials such as Ni-Zn and Ga-Cu, which exhibits good catalytic activity for the CO₂ to methanol conversion reaction [5, 6], validating our approach. Furthermore, we have discovered new potential materials, including Cu-Pt, which we intend to evaluate further in collaboration with our experimental team. This approach can be readily modified for the screening of potential heterocatalyst candidates when crucial reaction intermediates are known.<br/><br/>References:<br/>[1] L.-H. Mou <i>et al</i>., Adv. Sci. 10, 2301020 (2023)<br/>[2] J. K. Nørskov <i>et. al.,</i> J. Catal 209 275–278 (2002)<br/>[3] I L. Chanussot <i>et al.</i>, ACS Catal. 11 6059-6072 (2021)<br/>[4] G. Pacchioni, ACS Catal. 14 4 2730–2745 (2024)<br/>[5] P. Dongapure <i>et al.</i>, ChemCatChem 15, e202201150 (2023)<br/>[6] J. Zhong <i>et al</i>., J. Phys. Chem. C 125 2 1361–1367 (2021)