Description
Accurate prediction of excited-state energetics is essential for the rational design of organic semiconductors that enable next-generation optoelectronic technologies.(1) It is also a prerequisite for reliable machine-learning workflows, which depend on high-quality reference data to learn physically meaningful structure–property relationships.(2) Yet widely used single-reference methods, most notably TD-DFT, can break down for precisely the systems of greatest chemical and technological interest. These include molecules with pronounced static correlation and those exhibiting inverted singlet–triplet gaps (INVEST),(3) a key target property for high-efficiency OLED emitters and emerging photovoltaic concepts.(4)
We introduce the first multireference benchmark dataset tailored to small-molecule organic semiconductors. The collection comprises 1,500 chemically diverse π-conjugated compounds with systematically computed vertical excitation energies for S₁, S₂, T₁, and T₂, together with oscillator strengths for the lowest singlet transitions f₁ and f₂. All values are generated using a consistent multiconfigurational protocol based on state-averaged CASSCF and dynamical correlation via strongly contracted NEVPT2, providing a robust open-access reference for assessing excited-state methodologies on realistic materials motifs.
Beyond the dataset itself, we statistically interrogate method-dependent errors across the full dataset to assess the overall reliability of lower-cost single-reference approaches. We then extend this analysis to individual structural families, identifying cases where lower theory levels may remain sufficient even when dataset-wide trends indicate that multireference treatments are required. This provides practical guidance on when single-reference methods are dependable, when multireference descriptions become necessary, and which molecular families are most sensitive to the chosen level of theory.
Finally, we demonstrate the utility of the benchmark for materials discovery by screening for OLED-relevant INVEST, TADF, and anti-Kasha candidates. Strikingly, INVEST candidates are only identified at the NEVPT2 level, with all 21 molecules missed by both TD-DFT and CASSCF alone. For TADF and anti-Kasha screening, lower-cost methods recover some candidates, but 55 TADF and 82 anti-Kasha candidates remain NEVPT2-exclusive. These results show that dynamical correlation is indispensable for inverted-gap emitters and remains essential for reliable excited-state materials screening more broadly.
References
1. O. Ostroverkhova, Chemical Reviews, 2016, 116, 13279-13412
2. M. Zollner, Y. Moshfeghi, T. NematiaramDigital Discovery, 2026, 5, 1037-1067
3. L. Tučková, M. Straka R. R. Valiev and D. Sundholm, Physical Chemistry Chemical Physics, 2022, 24, 18713-18721
4. L. Barneschi, L. Rotondi, and D. Padula, Journal of Physical Chemistry A, 2024, 128, 2417−2426