A 2026 study by Liang and Head-Gordon (arXiv:2603.23466) argues, based on a systematic search of a functional space estimated to be 107610^{76} times larger than that of the previously best hybrid functional, that the rung-4 range-separated hybrid meta-GGA design space is approaching saturation, so that further systematic gains in accuracy and transferability for molecular ground-state chemistry will likely require going beyond the semi-local or hybrid paradigm. The problem is to construct a reproducibly specified, computationally practical density functional approximation that incorporates genuinely nonlocal exchange-correlation information — or, along the alternative routes the source identifies, a rung-5 double hybrid with its documented practical problems resolved, or a functional correcting strong-correlation (fractional-spin) error — and to demonstrate by broad, balanced molecular benchmarks that it surpasses the best rung-4 functionals (COACH, ωB97M-V) in accuracy while maintaining or improving transferability. The strongest existing claim in this direction (the Skala deep-learning functional, WTMAD-2 of 2.80 kcal/mol on GMTKN55 at semi-local cost) remains an unvalidated preprint whose headline claim is confined to main-group chemistry, that itself defers strongly correlated systems and covers transition metals only minimally, leaving the core open.

Background

Density functional theory (DFT) is the most widely used electronic-structure method in chemistry and materials science because it balances accuracy and computational cost. In its Kohn–Sham formulation, all many-electron effects are folded into the exchange-correlation (XC) functional, which must be approximated in practice. Approximate functionals are organized on Perdew's "Jacob's Ladder": rung 1 uses only the electron density (LDA), rung 2 adds the density gradient (GGA), rung 3 adds the kinetic-energy density τ\tau (meta-GGA, mGGA), rung 4 mixes in a fraction of exact Hartree–Fock exchange (hybrids; in range-separated hybrids, RSH, the mixing fraction varies with interelectronic separation), and rung 5 adds a second-order perturbative correlation term of MP2 type (double hybrids). The functional ωB97M-V (Mardirossian and Head-Gordon, 2016), a combinatorially optimized RSH meta-GGA with VV10 nonlocal van der Waals correlation (Vydrov and Van Voorhis, 2010 — a kernel depending on densities at two points, but only for dispersion), has been widely regarded as among the most accurate rung-4 functionals for molecular chemistry.

Liang and Head-Gordon (2026) developed a systematic design protocol — constraint enforcement, flexible functional forms, and modern optimization, trained on the balanced Gold-Standard Chemical Database 137 (GSCDB137; Liang and Head-Gordon, 2025), a diverse set of accurate energy differences for developing and testing functionals — and applied it within the RSH meta-GGA framework, producing the COACH functional, which improves accuracy and transferability over leading RSH meta-GGAs including ωB97M-V. Crucially, after exploring a functional space estimated to be 107610^{76} times larger than ωB97M-V's, they conclude that any further improvement within this framework will likely be smaller than their improvement over ωB97M-V — that is, the rung-4 semi-local/hybrid design space is approaching saturation. Their analysis attributes the remaining accuracy limits to delocalization error (failure to be exact for one-electron systems, i.e. spurious charge spreading) and to strong-correlation error — the fractional-spin/static-correlation breakdown when several electron configurations contribute comparably to the ground state; Cohen, Mori-Sánchez, and Yang (2008) proved that the exact functional must give a spin-interpolation energy that is constant for fractional-spin ensembles and that all approximate functionals violate this condition. For measuring broad accuracy, the community standard is GMTKN55, a database of 55 benchmark subsets (~1500 reactions) covering main-group thermochemistry, kinetics, and noncovalent interactions, summarized by the weighted total mean absolute deviation WTMAD-2 in kcal/mol.

The paper identifies three routes beyond rung 4. First, genuinely nonlocal XC functionals — using information beyond the semi-local variables and beyond fixed dispersion kernels like VV10 — are called "far less extensively explored" and promising; machine-learned functionals are a synergistic development, with the source citing Skala (Luise et al., 2025, arXiv:2506.14665), which learns nonlocal representations of electronic structure directly from data; the earlier DM21 (Kirkpatrick et al., 2021), which trained a neural-network functional under fractional-charge and fractional-spin constraints, is prior work of the same kind. Second, double hybrids (reviewed by Martin and Santra, 2020) already reach accuracies approaching wavefunction methods on GMTKN55 at hybrid-like cost when resolution-of-identity MP2 acceleration is available, but suffer from practical problems (slow MP2 basis-set convergence, frozen-core-protocol sensitivity, and spin contamination/spin-symmetry breaking — accuracy that is "impressive but not turnkey") that block their adoption as the next practical step. Third, strong correlation via DFT remains a frontier: multiconfiguration pair-density functional theory (Gagliardi et al., 2017) and fractional-spin corrections (Su, Li, and Yang, 2018) are promising but unintegrated with a transferable, practical ground-state functional. The authors also acknowledge secondary limitations — persistent grid sensitivity of some low-frequency noncovalent vibrational frequencies, and untested generalization to periodic metals, plane-wave implementations, magnetic response (current-DFT, the formulation of DFT in terms of the physical current density rather than the density alone), and excitations (time-dependent DFT) — which they place outside the scope of the central objective.

Problem Statement

How can density functional approximations for molecular ground-state chemistry be systematically improved beyond the demonstrated performance limit of the range-separated hybrid meta-GGA (rung 4) design space? Specifically, construct and validate a reproducibly specified, computationally practical exchange-correlation functional that incorporates genuinely nonlocal information (beyond semi-local density variables and fixed van der Waals kernels such as VV10) — or, along the alternative routes the source identifies, a rung-5 double hybrid with its documented practical problems (slow MP2 basis-set convergence / basis-set incompleteness, frozen-core-protocol sensitivity, and spin contamination / spin-symmetry breaking) resolved, or a functional correcting the fractional-spin/static-correlation error — such that, evaluated on broad and balanced molecular benchmark suites (GMTKN55- and GSCDB137-class) under community-standard protocols, it surpasses the best rung-4 functionals (COACH, ωB97M-V) in overall accuracy while maintaining or improving transferability, including on systems outside its training data and on strongly correlated (fractional-spin) and transition-metal-containing molecular cases. A complete answer comprises: the functional form and its parameters (or model weights) with full public specification; the design protocol by which it was obtained; and a demonstration, reproducible by benchmark replay, of the claimed superiority and of acceptable computational cost relative to hybrid DFT. A valid alternative answer form is a constructive refutation of the premise: a functional remaining strictly within the semi-local/hybrid paradigm that demonstrably surpasses COACH by a margin clearly exceeding the source's reported improvement over ωB97M-V, which would show that the rung-4 design space is not in fact saturated. The grid-sensitivity of low-frequency noncovalent vibrational frequencies and generalization to periodic metals, plane-wave implementations, magnetic response, and excited states are acknowledged secondary limitations of the source work, lying outside the paper's stated scope and therefore not part of this completion boundary.

References

  1. Liang, Jiashu; Head-Gordon, Martin (2026). Reaching for the performance limit of hybrid density functional theory for molecular chemistry. arXiv:2603.23466, DOI 10.48550/arxiv.2603.23466, https://arxiv.org/abs/2603.23466
  2. Liang, Jiashu; Head-Gordon, Martin (2025). Gold-Standard Chemical Database 137 (GSCDB137): A Diverse Set of Accurate Energy Differences for Assessing and Developing Density Functionals. Journal of Chemical Theory and Computation 21, 12601–12621, DOI 10.1021/acs.jctc.5c01380, https://doi.org/10.1021/acs.jctc.5c01380
  3. Mardirossian, Narbe; Head-Gordon, Martin (2016). ωB97M-V: A combinatorially optimized, range-separated hybrid, meta-GGA density functional with VV10 nonlocal correlation. The Journal of Chemical Physics 144, 214110, DOI 10.1063/1.4952647, https://doi.org/10.1063/1.4952647
  4. Vydrov, Oleg A.; Van Voorhis, Troy (2010). Nonlocal van der Waals density functional: the simpler the better. The Journal of Chemical Physics 133, 244103, DOI 10.1063/1.3521275, https://doi.org/10.1063/1.3521275
  5. Luise, Giulia; Huang, Chin-Wei; Vogels, Thijs; Kooi, Derk P.; Ehlert, Sebastian; Lanius, Stephanie; Giesbertz, Klaas J. H.; Karton, Amir; Gunceler, Deniz; Battaglia, Stefano; Simm, Gregor N. C.; Szabó, P. Bernát; Stanley, Megan; Bruinsma, Wessel P.; Huang, Lin; Wei, Xinran; Torres, José Garrido; Katbashev, Abylay; Zavaleta, Rodrigo Chavez; Máté, Bálint; Kaba, Sékou-Oumar; Sordillo, Roberto; Chen, Yingrong; Williams-Young, David B.; Bishop, Christopher M.; Hermann, Jan; Berg, Rianne van den; Gori-Giorgi, Paola (2025). Accurate and scalable exchange-correlation with deep learning. arXiv:2506.14665 (v6 revised 2026-04-21), DOI 10.48550/arXiv.2506.14665, https://arxiv.org/abs/2506.14665
  6. Wodyński, Artur; Kaupp, Martin (2026). Local-Hybrid Functional With a Composite Local Mixing Function Built From a Neural Network and a Strong-Correlation Model. Journal of Computational Chemistry 47, e70294, DOI 10.1002/jcc.70294, https://doi.org/10.1002/jcc.70294
  7. Martin, Jan M. L.; Santra, Golokeshwar (2020). Empirical Double-Hybrid Density Functional Theory: A 'Third Way' in Between WFT and DFT. Israel Journal of Chemistry 60, 787–804, DOI 10.1002/ijch.201900114, https://doi.org/10.1002/ijch.201900114
  8. Cohen, Aron J.; Mori-Sánchez, Paula; Yang, Weitao (2008). Fractional spins and static correlation error in density functional theory. The Journal of Chemical Physics 129, 121104, DOI 10.1063/1.2987202, https://doi.org/10.1063/1.2987202
  9. Kirkpatrick, James; McMorrow, Brendan; Turban, David H. P.; Gaunt, Alexander L.; Spencer, James S.; Matthews, Alexander G. D. G.; Obika, Annette; Thiry, Louis; Fortunato, Meire; Pfau, David; Castellanos, Lara Román; Petersen, Stig; Nelson, Alexander W. R.; Kohli, Pushmeet; Mori-Sánchez, Paula; Hassabis, Demis; Cohen, Aron J. (2021). Pushing the frontiers of density functionals by solving the fractional electron problem. Science 374, 1385–1389, DOI 10.1126/science.abj6511, https://doi.org/10.1126/science.abj6511
  10. Gagliardi, Laura; Truhlar, Donald G.; Li Manni, Giovanni; Carlson, Rebecca K.; Hoyer, Courtney E.; Bao, Junwei Lucas (2017). Multiconfiguration Pair-Density Functional Theory: A New Way To Treat Strongly Correlated Systems. Accounts of Chemical Research 50, 66–73, DOI 10.1021/acs.accounts.6b00471, https://doi.org/10.1021/acs.accounts.6b00471
  11. Su, Neil Q.; Li, Chen; Yang, Weitao (2018). Describing strong correlation with fractional-spin correction in density functional theory. Proceedings of the National Academy of Sciences 115, 9678–9683, DOI 10.1073/pnas.1807095115, https://doi.org/10.1073/pnas.1807095115