OSOS brings radical chemistry into view with a benchmark built for doublet systems
The Open-Shell Organic Systems database arrives as a corrective to a long-standing blind spot in large computational datasets. For years, benchmark collections such as QM9 and ANI-1x have leaned heavily on closed-shell molecules, leaving open-shell organic radicals underrepresented despite their central role in combustion, atmospheric chemistry, polymer degradation, and many reaction networks used to train modern computational models. OSOS fills that gap with 9,934 reaction energies calculated at the CCSD(T)/CBS limit through W1-F12 theory, giving method developers a stringent reference set for doublet radicals rather than a convenience sample of easy cases.
The database spans localized and delocalized radicals built from carbon, nitrogen, oxygen, and fluorine, with chemical space limited to structures containing up to seven non-hydrogen atoms. Its structure is deliberate rather than incidental. Radical chemistry is notoriously sensitive to electron correlation, spin distribution, and resonance effects, and those features often expose weaknesses in approximate methods long before they fail on more benign closed-shell systems.
One sentence captures the problem plainly: most standard density functionals do not treat open-shell reaction energetics kindly.
A dataset shaped around atomization, abstraction, shift, and isomerization chemistry
OSOS contains 3,360 radical total atomization energies, 3,360 H-abstractions, 2,292 H-shifts, and 922 constitutional isomerizations. Each category probes a different aspect of electronic structure theory. Atomization energies test the balance between bond breaking and formation, H-abstractions challenge methods to capture bond dissociation and radical stabilization, H-shifts force attention to intramolecular rearrangements, and isomerizations examine whether a method can preserve relative stability across distinct constitutional forms.
That mix matters because radical chemistry does not fail in one uniform way. A functional can look acceptable on a narrow class of localized species and then drift sharply when electron density is delocalized over multiple atoms. The OSOS design makes those weaknesses visible instead of hiding them in aggregate averages.
The dataset also closes a practical modeling gap for machine-learning potentials. High-quality reference energies are the fuel these models require, and the quality of the training signal becomes especially important when the target systems are electronically demanding. A model trained mainly on closed-shell species may interpolate smoothly there while remaining poorly calibrated for radical intermediates, which are often the species of greatest mechanistic interest.
Density functional theory overestimates radical atomization energies across most methods
The comparative analysis of density functional theory shows a consistent pattern: most functionals systematically overestimate radical atomization energies. The direction of error is not random noise but a bias, which is more troubling for method users because it implies a structural weakness in the approximations themselves. In practical terms, overestimating atomization energies can distort predicted reaction thermochemistry, misplace relative stabilities, and alter conclusions about which pathways are favorable.
Dispersion corrections do not rescue the problem here. In this radical benchmark, adding dispersion generally worsens performance rather than improving it. For a field accustomed to treating dispersion as a broadly helpful correction, this is a sharp reminder that correction schemes are not universally beneficial. When the dominant challenge is electronic structure in open-shell species, bolting on an empirical dispersion term may not address the main source of error and can even push results farther from the reference values.
The article also separates radicals by electronic character, and the contrast is striking. Delocalized radicals are much harder for most DFT methods than localized species, producing larger overestimations of atomization energies. In localized radicals, the error pattern rises in the order O-centered < C-centered < N-centered. That progression suggests that the underlying error depends not just on whether a radical is open-shell, but on where the unpaired electron resides and how strongly the local environment perturbs its distribution.
Why delocalization strains approximate methods
Delocalized radicals are difficult because the unpaired electron is spread across a broader framework, which amplifies sensitivity to self-interaction error, exchange balance, and the treatment of near-degeneracy. Approximate functionals often handle one-electron localization and multi-center electron sharing differently, and those differences become visible when resonance stabilization is part of the energetic balance. Localized radicals, by contrast, present a narrower electronic problem, although the identity of the heteroatom still changes the level of difficulty.
The O-, C-, and N-centered ranking for localized species likely reflects how each atomic center interacts with surrounding substituents and how strongly the method must cope with polarization and spin density placement. Nitrogen-centered radicals appear to be the most difficult of the localized set in this benchmark, while oxygen-centered radicals are the least problematic of the three. In a database built for method stress-testing, these differences are useful because they give developers a route to diagnose where a functional is failing rather than merely observing that it is failing.
Skala stands out with sub-kcal mol-1 performance across reaction classes
Against this difficult backdrop, the deep-learning meta-GGA Skala performs exceptionally well. It outperforms conventional functionals across all rungs of Jacob’s Ladder and consistently achieves mean absolute deviations below 1 kcal mol-1 for abstraction, shift, and isomerization reactions. For a benchmark centered on open-shell radical chemistry, that level of agreement is notable because it suggests the method can reproduce subtle energetic differences without collapsing under the complexity of radical electronic structure.
The result is especially important because it points to a path beyond incremental functional tuning. A model that maintains sub-kcal mol-1 accuracy across multiple radical reaction types is not just fitting a single benchmark slice; it is addressing the broader challenge of open-shell reactivity with a level of robustness that conventional functionals do not match in this dataset.
Still, even a strong result within one benchmark should be read as a test of current capability rather than proof of general mastery. The OSOS collection is deliberately focused on organic doublet radicals containing only C, N, O, and F, and it spans systems with up to seven non-hydrogen atoms. That scope is rigorous, but it is still bounded. Larger radicals, other elemental compositions, and more complex environments will need their own validation before anyone can assume equivalent performance there.
A reference set built to reset expectations for radical modeling
OSOS sets a higher bar for both density functional development and machine-learning potential training. By anchoring evaluation to CCSD(T)/CBS-quality reaction energies for doublet radicals, it gives researchers a way to test whether a method is genuinely learning radical chemistry or merely succeeding on the easier closed-shell cases that have long dominated public datasets.
The most useful feature of the database is not just its size, but its selectivity. It gathers a chemically diverse set of open-shell organic systems and organizes them around reaction classes that probe distinct electronic demands. That design makes it a practical benchmark for identifying systematic bias, not simply a large archive of numbers.
For method developers, the immediate challenge is clear: any new functional or machine-learning potential claiming broad reliability now has to survive on a dataset built expressly to expose failures in radical chemistry. The unresolved question is how far these gains will hold when the same methods are pushed into broader chemical space, where open-shell species interact with larger structures, more elements, and more complicated energetic landscapes.