Toward AI/ML-assisted Discovery of Transition Metal Complexes

Traditional computational methods for molecule design are based on first principles calculation, which places a high demand on computing power. The increasingly powerful machine learning (ML) models have fundamentally transformed this landscape. Statistically, by learning the joint probability distribution between molecular or material structure and targeted properties, generative models can autonomously design numerous novel structures with satisfactory properties. This inverse design strategy clearly outperforms the traditional physics-based methods which requires human expertise and intuition, along with serendipity. To validate the generated molecules or materials for specific properties, classical discriminative models allow for fast large-scale screening of the quantitative structure-activity relationships. Generally, the completely ML-based workflow from generation to validation for the exploration of chemical space is accessible and provides outstanding benefits which traditional computational approaches struggle to achieve. In this review, we summarize recent advances in ML-assisted discovery for transition metal complexes and conclude with several existing challenges which impede the widespread practical applications of this technology to the class of problems.