Computational data as the fuel for AI in chemistry

The rapid growth of machine learning and generative AI in the chemical sciences has placed increasing demands on the data used to train and evaluate these models. Although algorithmic advances have attracted considerable attention, the quality, consistency, coverage and accessibility of chemical data remain major constraints on AI-driven discovery. This perspective argues that computational data should be regarded not merely as a complement to experimental data, but as deliberately designed, domain-specific infrastructure for chemical AI. This does not require exhaustive coverage of chemical space: the distinctive value of computational methods lies in their ability to generate reproducible, consistently labelled data systematically within selected regions of chemical and configurational space. Through case studies in neural-network potentials and computational peptide design, we examine how the selection, sampling, fidelity and scale of computational datasets define model capabilities and domains of applicability. We then consider how computational and experimental data can be combined, with computation enabling scalable, targeted data generation and experiment providing physical grounding and validation. We conclude by identifying priorities for deliberate dataset design, FAIR data practices and open benchmarking in chemical AI.