Defining chemical problems for artificial intelligence models

Abstract

Artificial intelligence has become an increasingly important tool in chemistry, yet its impact is often constrained by how chemical problems are defined for machine learning. In this Personal Account, I argue that dataset construction plays a central role in AI-driven chemistry, not merely as a preparatory step but as a means of formalizing new chemical learning problems. I summarize recent efforts to build machine-learning-ready datasets in three domains: molecular property prediction, property-conditioned molecular design, and reaction mechanism prediction. Experimentally measured optical property datasets were curated to explicitly incorporate environmental context, enabling models to learn structure–environment–property relationships that reflect real experimental conditions and to support inverse design of functional chromophores. For chemical reactivity, large-scale mechanistic datasets were constructed through rule-based imputation of elementary steps, transforming overall reactions into supervised learning problems for mechanism prediction. Limitations of sequence-based reaction models motivated the development of FlowER, which represents reactions as electron redistribution processes using bond–electron matrices and enforces mass conservation by construction, allowing direct integration with quantum chemical validation. Together, these studies illustrate a unifying perspective in which datasets, representations, models, and downstream workflows are designed jointly, guided by chemical intuition. By framing dataset construction as an act of problem definition, this Account highlights how chemists can shape the scope and direction of artificial intelligence as a tool for chemical discovery.