Over the past two decades, the integration of machine learning (ML) into theoretical and computational chemistry has transformed the scale and scope of discovery that is possible on a computer. In this Perspective, I share my personal journey from early density functional theory method development for small transition metal complexes to the building of automated workflows and ML models for catalysis, redox chemistry, and materials design. I will describe my unlikely path toward machine learning starting from minimal models of catalysts and the use of structural databases to gain data-driven insights. I will then describe our experiences in training machine learning models for discovery, including active learning and descriptor-based approaches that enabled data-driven exploration despite limited experimental reference data for open-shell transition metal complexes. As an example of overcoming these limitations, I will describe our more recent efforts that culminated in experimental validation of our computational predictions. Along the way, challenges in data curation, DFT method sensitivity, and synthetic realism have shaped the trajectory of the field. I conclude with reflections on the rapid rise of generative AI, agentic workflows, and the enduring need to ask questions that truly advance molecular discovery.
Heather J. Kulik (Wed,) studied this question.