The Open-Shell Organic Systems Database: Diverse CCSD(T)/CBS Reaction Energies for Assessing Density Functional Theory and Machine Learning Methods

Abstract

Abstract We present the Open-Shell Organic Systems (OSOS) database, a diverse benchmark set comprising 9934 atomization, isomerization, and H-abstraction energies computed at the CCSD(T)/CBS limit via W1-F12 theory. Historically, large-scale data sets (e.g., QM9 and ANI-1x) have been heavily biased toward closed-shell species. The OSOS database addresses this gap by providing a diverse and accurate set of reaction energies dedicated exclusively to organic doublet radical systems for benchmarking and training purposes. The data set covers the chemical space spanned by localized and delocalized organic radicals containing up to seven non-hydrogen atoms (C, N, O, and F). The database consists of 3360 radical total atomization energies, 3360 H-abstractions, 2292 H-shifts, and 922 constitutional isomerizations. We use this database to assess the performance of density functional theory (DFT). We find that most density functionals systematically overestimate radical atomization energies; therefore, the inclusion of dispersion corrections generally leads to a deterioration in performance. Furthermore, delocalized radicals present a significantly greater challenge for most DFT methods than localized species, leading to larger overestimations of their atomization energies. For localized radicals, errors increase in the order: O-centered < C-centered < N-centered. Remarkably, the deep-learning meta-GGA Skala functional demonstrates exceptional accuracy, significantly outperforming conventional functionals across all rungs of Jacob’s Ladder and consistently achieving mean absolute deviations below 1 kcal mol–1 for the abstraction, shift, and isomerization reactions. The OSOS database establishes a rigorous standard to guide the development and validation of next-generation density functionals and machine-learning potentials for doublet radicals.