Design and assembly of virtual homogeneous catalyst libraries - towards in silico catalyst optimisation

An alternative top-down concept for searching for homogeneous catalysts is introduced. In this approach, three multi-dimensional spaces are considered. These represent the catalysts, the descriptor values (e.g., cone angle, lipophilicity indices), and the figures of merit (e.g., turnover frequency, enantiomeric excess, or product selectivity), respectively. By generating and connecting these spaces, it is possible to screen virtual catalyst libraries and indicate regions in the catalyst space where good catalysts are likely to be found. The generation of the catalyst space from simple building blocks is presented and the application of this approach is demonstrated for two cases: predicting the bite angle in a library of 600 ligand-Rh complexes, and predicting the linear : branched aldehyde product ratio that these 600 catalysts would give in the hydroformylation of 1-octene. In the latter case, the model is first trained on a set of 39 octene hydroformylation reactions. The limitations and applications of this concept are discussed.