While machine learning (ML) methods are already commonly applied in heterogeneous catalysis, the use of such methods for the design of homogeneous catalysts is a largely overlooked field. A recent proof-of-principle study showed the huge potential of ML in homogeneous catalysis by demonstrating that activation barriers in a set of related transition metal (TM) complexes can be learned. ML4Catalysis has three objectives that go far beyond this state of the art: 1) Automation of quantum chemistry (QC) calculations by combining different existing computational tools in a unified framework, with the goal to create powerful high-level computational workflows in a synergistic way. 2) Going beyond the accuracy of density-functional theory (DFT), which is often inaccurate for systems with multireference (MR) character like TM complexes. To this end, we will develop an ML method that is trained to predict the difference between energies at the DFT level and at a more accurate multireference level. 3) A pool of entirely novel catalysts for a given reaction will be generated by using a variational autoencoder (VAE) architecture. A Gaussian Process (GP) model trained to predict key activation barriers on a subset of these complexes will be used to screen the remaining set for the most promising candidates. This approach will be applied to find novel CO2 hydrogenation catalysts, which are important for the creation of fuels and feedstock chemicals from natural resources. With its focus on catalysis and modern QC and ML methods, ML4Catalysis is highly relevant for two of the European Commission’s current priorities: “A European Green Deal” and “A Europe fit for the digital age”. The interdisciplinary project combines knowledge of the researcher on modern MR methods and the electronic structure of TM complexes with the expertise in automation, ML, and homogeneous catalysis at the host institution and will leave the researcher well-prepared for an independent career.