Benchmarking electronic structure methods using generative AI to construct diverse reaction databases

Highly diverse chemical databases play a key role in the development of accurate and generally applicable electronic structure methods. We develop and test a workflow for autonomously generating reaction databases using large language models. As a proof-of-concept, we generate a reaction database encompassing Diels–Alder cycloaddition, electrocyclic ring opening, electrophilic addition, electrophilic aromatic substitution, and acid-base reactions. We proceed to benchmark DFT methods relative to CCSD(T) reaction energies obtained via G4(MP2) theory. The best-performing DFT methods from each of the rungs of Jacob's Ladder (mean-absolute-deviations in parentheses) are: PBE-D4 (1.91), Skala (1.86), CAM-B3LYP-D4 and ωB97X-D (1.49), and B1B95-D4 (1.48 kcal mol−1).