Improving coupled cluster theory for strongly correlated molecules with Kohn-Sham density encoding Authors: Abdulrahman Y Zamani, Barbaro Zulueta, Andrew M Ricciuti, John A Keith, Kay Carter-Fenk Published: 2026 ABSTRACT AbstractCoupled cluster theory using a Kohn-Sham density functional reference can dramatically outperform the standard Hartree-Fock-based approach for strongly correlated systems, but the origin of these improvements has remained unclear. We show these improvements arise from differences in the one-particle density as encoded into the Fock matrix, not from the nature of the orbitals themselves, as is commonly assumed. Exploiting this insight, Kohn-Sham references enable near-chemical accuracy for electronic and thermochemical properties of transition metal dimers and main group compounds. Notably, Kohn-Sham coupled cluster qualitatively recovers the entire chromium dimer potential energy surface, a notorious failure case for conventional methods. We further introduce a density difference diagnostic that detects multireference character and guides the selection of optimal references at mean-field cost. Together, these results establish a practical route to treating strong correlation within quantum chemistry’s “gold standard” framework, with immediate implications for machine learning potential development and materials research, which rely heavily on density functional theory.Subject terms: Quantum chemistry, Method development, Density functional theory Introduction Chemical catalysis by transition metal compounds (TMCs) underpins renewable energy technologies1–5, yet predicting their thermochemistry and kinetics from first principles remains challenging. Heightened interest in replacing precious metals with Earth-abundant alternatives6 amplifies this challenge since the same first-row transition metals prized for sustainability (e.g., iron, cobalt, nickel, copper, manganese) often exhibit the multiconfigurational electronic structures that confound standard quantum-chemical methods. Coupled cluster (CC) theory with singles, doubles, and perturbative triples [CCSD(T)]7,8 achieves “gold standard" accuracy for main group thermochemistry, but its reliability for TMCs can be compromised by strong correlation arising from the d-orbital manifold9–11. This limitation has profound consequences: CCSD(T) reference data train machine learning potentials12 and parameterize semi-empirical methods13,14 that enable exascale simulations15–17, yet systematic errors for TMCs propagate through this entire computational ecosystem. Replacing the Hartree-Fock (HF) reference with one from Kohn-Sham density functional theory18,19 (KS-DFT) offers a tantalizing solution. KS-CC methods have yielded accurate results for specific transition metal thermochemistry20,21, spin-state energetics22–25, barrier heights, and reaction kinetics26,27. However, improvements over HF-CC are maddeningly inconsistent: comprehensive assessments reveal no systematic advantage, and the choice of density functional has minimal impact when improvements do occur28. The prevailing explanation has been that KS orbitals resemble Brueckner orbitals29–31 and therefore accelerate the convergence of truncated CC to the full configuration interaction limit, but this explanation fails to account for the fact that the KS orbitals are not actually the orbitals used in a KS-CC calculation. Orbital-based arguments are therefore inconsistent with the reality of practical KS-CC calculations and thus fail to account for these erratic outcomes 32,33. The present work resolves this puzzle. We demonstrate that improvements from KS-CC arise not from the orbitals themselves, but actually from differences in the one-particle density matrix that are encoded into the non-canonical Fock operator34. In practical KS-CC calculations, the KS density is iterated to self-consistency, then the Hartree-Fock potential is used to build the one-electron Fock matrix, and the orbitals are semi-canonicalized to ensure that their energies are meaningful. This procedure has the critical consequence of transforming the KS orbitals toward the HF result while preserving the density, meaning the orbitals that enter CC calculations bear little resemblance to the converged KS solution. Thus, the KS density matrix itself acts as the operative quantity by etching information about electron correlation into the Fock matrix. This Fock matrix is then used to compute the CC wavefunction and, under the right conditions, can drive the CC equations toward a more accurate solution. With this insight, we systematically assess KS-CCSD(T) for first-row transition-metal diatomics (M-O, M-Cl, M-H+, M-H, and M-M; where M = Sc-Zn)35–37 and main-group species, benchmarking against experiment and higher-level wavefunction methods across Rungs 1–4 of Jacob’s ladder38. We achieve substantial improvements for systems with purported multireference character39–41. Most notably, KS-CC qualitatively recovers the entire Cr2 potential energy surface, a notorious failure case normally assumed to require multireference methods since methods such as HF-CCSD(T) and hybrid DFT functionals generally perform less well than their lower-rung counterparts42,43. Finally, we propose a density difference metric that introduces a new multireference diagnostic at mean-field cost and provides a rational basis for selecting optimal reference densities in KS-CC calculations. Results Preconditioning coupled cluster with Kohn-Sham densities To identify the physical origins of differences in CC results when beginning with a KS solution, we first examine the effect of the semi-canonicalization44–46 of initial reference single-particle quantities that are typically used in post-self-consistent-field (SCF) implementations. Semi-canonicalization involves diagonalizing the occupied-occupied and virtual-virtual blocks of the KS Fock matrix F~, constructed from a converged KS density and the Fock operator F. The resulting unitary transformation is applied to the KS orbitals, and a new set of single-particle energies is provided. This transformation does not generally result in zero entries for occupied-virtual elements, F~ov. Generalized CC implementations can accommodate arbitrary single-reference determinants, but when the Fock operator is not strictly diagonal, obtaining physically meaningful orbital energies and MOs requires a semi-canonicalization step. This is especially important for KS-CC calculations, where the physicality of KS eigenvalues remains ambiguous at best47,48 unless recast into a correlated orbital theory49–56. When using KS references in correlated wave function calculations, the F~ov elements (denoted fia below) carry essential information about the KS density, appearing as non-Brillouin singles (NBS) corrections in many-body perturbation theory (MBPT), and in terms that depend on elements of F~ (fpq) in the CC amplitude equations57,58 where tpq and tprqs are single and double substitution amplitudes, respectively. We illustrate this point by computing molecular orbital (MO) energies for N2 with HF and density functional approximations (DFAs) that contain different amounts of exact Hartree-Fock exchange (HFX). Fig. 1 shows that upon semi-canonicalization of the KS orbitals, their energies will closely resemble the HF eigenvalues. This effect remains stable with respect to increases in %HFX. Conversely, canonical occupied KS eigenvalues are vastly different from those of HF, with the differences between the two vanishing as more HFX is included in the DFA. The semi-canonicalization effect appears even stronger on virtual MOs; the semi-canonical “KS" virtual MO energies align with their HF counterparts and are stable against increasing %HFX. Larger molecules exhibit similar behavior when semi-empirical functionals are employed (see Supplementary Table 8). MO eigenvalue changes with increasing %HFX for HF, B3LYP, and B3LYP after semi-canonicalization. The percent exchange of the DFA is adjusted in proportion to the exact exchange. The vertical dashed lines represent the standard %HFX in the DFA. Highest occupied (HOMO) and lowest unoccupied (LUMO) MO energies in Hartree (Ha) are obtained with the cc-pVTZ and aug-cc-pVTZ basis sets, respectively. Modifying the single-particle quantities significantly affects the energy denominator magnitudes in CC and Møller-Plesset (MP) theory amplitude expressions (see Supplementary Figs. 46, 47), including those with NBS contributions that would otherwise vanish for a canonical HF reference. The eigenvectors after semi-canonicalization are also involved in the transformation of the two-electron integrals into the semi-canonical basis: ⟨pq∣∣rs⟩~59. We also confirm that semi-canonical orbital energies (ϵp~) become similar (but not necessarily identical) to the canonical HF eigenvalues. This suggests that the initial MOs and their associated energies are not the sole quantities responsible for changes in correlated interaction terms of post-SCF methods. Instead, the semi-canonicalization procedure exactly preserves the KS density matrix, which is invariant under unitary transformations, and encodes this density information into the Fock matrix elements that enter the MBPT/NBS and CC equations. Such is the case for any unitary transformation of the occupied or virtual KS MOs (i.e., MO localization); thus, the space the MOs span—encoded in the density—is the most pertinent factor in preconditioning the CC equations toward different solutions. Response properties using KS references Previous work leveraged semi-canonical KS references with Green’s function (GF) methods to obtain correlated orbital energies60. More recently, the renormalized singles (RS) approach for one- and two-shot GW methods adopted this step61–63. These methods incorporate essential NBS terms and yield accurate ionization potentials (IPs), electron affinities (EAs), and dipole moments. Benchmark studies show that RS GW methods with generalized gradient approximation (GGA) or hybrid functional reference densities outperform their HF-based counterparts. Here, we observe how KS densities influence diagonal 2nd and 3rd order self-energy approximations that are derived from electron propagator theory (EPT) for IPs and EAs (Supplementary Fig. 7 and Table 7). The semi-canonicalized eigenvalues obtained under the HF potential ϵp~ either closely restore64 or slightly improve Koopmans’ theorem for IPs. However, MAEs of self-energy-corrected orbital energies from KS-EPT deviate from HF-EPT by up to ~ 0.17 eV. An exception arises for EAs computed with EPT@B3LYP for BeO, C2, and BN—molecules noted to contain appreciable multireference (MR) character. Clearly, low-order perturbation theory does not fully alleviate the dependence of EPT methods on the initial single-determinant reference. Thus, methods that optimize or renormalize the reference state65,66, such as CC, can be beneficial for non-self-consistent EPT approaches 67,68. Finally, we test reference-state dependence for response properties by computing single excitation energies with equation-of-motion CC (EOM-CCSD) on the QUEST #1 dataset (see Supplementary Fig. 6)69–71. These molecules are presumed to be dominated by the initial (i.e., single reference) ground-state configuration. Here, semi-canonicalized KS references approach closely or reproduce HF-based results, and this indicates that CC solutions are qualitatively the same. Bond dissociation energies Having established how semi-canonicalization transforms single-particle quantities, we now assess whether KS-CC improves total energy differences for strongly correlated metal systems9,72–76. We benchmark KS-DFT, HF-CCSD(T), and KS-CCSD(T) for bond dissociation energies (BDEs) across transition metal diatomics. Figure 2 compares error measures against phaseless auxiliary field quantum Monte Carlo (ph-AFQMC) for M-H, M-O, and M-Cl species. CCSD(T) with optimally selected DFA or HF references achieves ph-AFQMC accuracy across all bond types (MAE = 0.07 eV vs. 0.05 eV for ph-AFQMC). KS-CCSD(T) outperforms standalone KS-DFT for these BDEs; in particular, ωB97X-V, ωB97M-V, and B3LYP references yield KS-CCSD(T) results that match or slightly surpass HF-CC accuracy. For M-H+ and M-M species (Fig. 3), the differences widen. Conventional complete basis set limit (CBS) CCSD(T) calculations with HF density fail dramatically for metal–metal bonds (MAE = 0.95 eV) and perform no better than the local (spin) density approximation (LDA) functional. However, CCSD(T)/CBS with GGA references (PBE, PW91) approaches chemical accuracy (MAE ≈ 0.11–0.14 eV) 77,78. Benchmark comparison of density functional and coupled cluster methods for transition metal–ligand bond dissociation energies. Signed errors in calculated bond dissociation energies, ΔDe=Decalc−Deexp, for first row transition metal compounds across four bond classes: a metal–oxygen (M-O); b metal–chlorine (M-Cl); c metal–hydrogen (M-H); and d over all bonds. Boxes span the interquartile range with the median marked, whiskers extend to 1.5 times the interquartile range, and circles show each individual species. White diamonds mark the mean absolute error (MAE). Blue boxes represent density functional methods (KS-DFT), orange boxes represent CCSD(T)/CBS calculations using different reference determinants (CC-X denotes CCSD(T)/CBS with densities from method X), and the green box represents phaseless auxiliary-field quantum Monte Carlo (ph-AFQMC) results from ref. 75. CC-Best denotes the optimal reference selection either from HF or DFT for each system, and QMC-Best denotes the optimal reference when using MP2, CC, or TZ/QZ extrapolation in ph-AFQMC. The gray shading indicates the chemical accuracy range77 ( ± 3 kcal/mol ≈ ± 0.13 eV). Performance of density functional and coupled cluster methods for metal hydride cations and bimetallic bond dissociation energies. Signed errors in calculated bond dissociation energies, ΔDe=Decalc−Deexp, for: a metal hydride cations (M-H+); b homonuclear metal--metal dimers (M-M); and c overall performance. Boxes span the interquartile range with the median marked, whiskers extend to 1.5 times the interquartile range, and circles show each individual species. White diamonds mark the mean absolute error (MAE). Blue boxes represent density functional methods (KS-DFT), and orange boxes represent CCSD(T)/CBS with different reference determinants. CC-Best denotes the optimal reference selection (HF or DFT) for each system. The gray shading indicates the chemical accuracy range77 ( ± 3 kcal/mol ≈ ± 0.13 eV). Accurately modeling reactivity and thermochemistry requires accurate potential energy curves (PECs). We therefore turn to the infamous Cr279,80 as a stringent test for our approach. Previous work applied KS-DFT and HF-CC independently to Cr2, finding that certain functionals yield qualitatively reliable PECs and accurate req81–85. Such success stories have also been born out for other qualitative descriptions of transition metal bonding, albeit strategies to guide the optimal choice of DFA remain unclear86–89. Curiously, Bauschlicher & Partridge90 showed that at req, UCCSD(T) underbinds, RCCSD(T) overbinds, and UBLYP provides the best comparison to experiment. Figure 4 shows PECs from KS-UCCSD(T) with spin-symmetry constraints removed to describe the antiferromagnetic state alongside a best theoretical estimate (BTE) and experimental data. Again, conventional UCCSD(T)@HF/CBS produces a qualitatively incorrect potential that has a shallow minimum shifted to longer bond lengths. Potential energy curve of the antiferromagnetic state of Cr2 dimer computed with UCCSD(T)/CBS using different reference densities. Binding energy as a function of bond length for Cr2, the prototypical multireference transition metal dimer. Plot depicts UCCSD(T)/CBS calculations with reference orbitals from Hartree-Fock (UCC-HF), SVWN5 (UCC-SVWN5), PBE (UCC-PBE), PW91 (UCC-PW91), R2SCAN (UCC-R2SCAN), and PBE0 (UCC-PBE0). The experimental curve (Exp, black) and the best theoretical estimate from state-of-the-art multireference calculations (BTE, crimson) from ref.79. provide comparison. KS-UCCSD(T)/CBS with PW91 or PBE qualitatively reproduces the PEC topography of both experiment and high-level MR theory, particularly in the characteristic shelf region. We find that KS-UCCSD(T) with these DFAs attains greater accuracy than HF-UCCSD(T) and even rivals the accuracy of sophisticated MR methods79,91–93 at much lower cost. We also observe that both DFAs deliver chemically accurate BDEs consistent with experiment94–96. Thus, the decades-long failure of CCSD(T) for Cr2 shown in refs. 79,90,97, is not a fundamental limitation of single-reference CC, as it can largely be remedied by simply employing alternatives to the canonical HF reference density. PBE and PW91 references reduce the UCCSD(T)/CBS errors from 0.52 to 0.001 eV for De and from 0.70 to 0.04 Å for req (Supplementary Table 6). Spin-state energetics Functional choice and exact exchange content significantly impact KS-DFT spin-state energetics98–103. Systematically improvable methods, such as CC should provide a more consistent route to accurate spin-dependent properties. We therefore examine singlet-triplet gaps for two classes of species: neutral transition metal diatomics (denoted as the TinySpins25 set) and main group molecules with established multireference character. Supplementary Notes 14 and 15 provide error measures, Frobenius norms of F~ov, T1 diagnostics, ⟨S2⟩ values, and additional qualitative analysis. By comparing singlet-triplet gaps computed with CCSD(T)/def2-QZVPPD against CBS HF-CCSDT(Q)Λ estimates, we find that hybrid functionals, notably SCAN0, closely reproduce HF-CC results (Supplementary Fig. 15). LDA and (meta-)GGA references perform less well but remain within 0.25 kcal/mol of the HF-CCSD(T) MAE. For the TinySpins25 set, KS densities bring no added benefit. Average T1diag values for singlet and triplet states exceed the heuristic threshold104 ( ~ 0.02), suggesting MR character. However, standard amplitude-based diagnostics may inadequately categorize molecules with pronounced nondynamical correlation; alternative thresholds (e.g. T1diag>0.05) have been proposed for inorganic systems39. Consistent with previous analyses20,28,105,106, KS references suppress singlet T1diag values below HF-CC thresholds. For triplets, only hybrid DFAs produce this suppression. We return to these observations in the section on classifying MR character. Next, we assess KS-CC performance on singlet-triplet gaps using main group molecules that higher-level theories predict to have MR ground states. Specifically, we consider an isoelectronic series (BN, C2, BO+, CN+) and a diradical series (CH2, O2, NF, NH). Fig. 5 compares singlet-triplet gaps for two prototypical MR species against experiment and high-level theoretical estimates, and Supplementary Fig. 25 presents the remaining species. BN and CN+ show clear improvements over HF-CC with various DFA references. Notably, HF-CC predicts a negative singlet-triplet gap for BN, which implies an incorrect ground-state assignment, while KS-CC correctly identifies the singlet ground state, consistent with Li and Paldus107. However, CH2, whose closed-shell singlet is partially MR108,109, shows no improvement with KS-CC. These examples illustrate both the promise and the limits of KS-CC for states with pronounced MR character. Singlet-triplet gaps for BN and CH2 computed with CCSD(T)/cc-pVTZ with HF and KS-DFT determinants. The solid and dashed horizontal orange line represent the reference theoretical estimate and experimental value, respectively107,108. Classifying multireference character KS references can treat some strong correlation effects110–112 that arise from MR character, but T1diag alone cannot predict whether KS-CC will improve results (Ref28, see Supplementary Fig. 48). Many other MR diagnostics exist106,113–119, but none reliably predicts KS-CC performance. Comparing KS-CC densities to near-full configuration interaction (FCI) densities via distance or entropic metrics could gauge reference quality, but generating CI-quality densities incurs substantial computational cost. Moreover, Benedek et al28. showed that KS-CC density errors relative to approximate FCI are more inconsistent than HF-CC errors. Overall, most metrics require the calculation of a wavefunction (CC or multi-configurational), which is suboptimal for the user who is interested in screening DFAs for their potential to improve CC calculations. It would therefore be highly desirable to define a proactive metric that can predict MR character before a computationally demanding wavefunction calculation is run. Our results imply that the KS density matrix used to construct F~ must differ sufficiently from its HF counterpart for KS-CC to yield a different solution than HF-CC. This raises a practical question: can we determine a priori at the SCF level whether a given DFA will produce significantly different CC results? Drawing on recent interpretive tools for ΔSCF methods, we develop a difference-density analysis based on the natural orbital decomposition of ΔP to quantify the charge density shift between HF and KS-DFT (see Methods). From this we define the normalized natural-orbital eigenvalue diagnostic (NNED) between SCF solutions. Fig. 6 illustrates this analysis for Cr2, supplemented by Frobenius norms ∣∣ΔP∣∣F. Supplementary Fig. 14 presents main group species. The NNED serves as an economical proxy for T1diag, capturing the magnitude of reference density changes. Applying NNED to assess MR character, we obtain classification thresholds proportionate to HF-CC T1diag limits. Supplementary Note 12 details the classification protocols. When NNED approaches or exceeds a T1-equivalent threshold, this suggests an MR treatment may be needed, and so testing alternative DFAs for improved CC results is warranted. Difference density natural orbitals for singlet Cr2 computed between HF and PW91 with def2-QZVPP. The six largest electron displacement eigenpairs δ± are reported. The orbitals corresponding to δ− and δ+ are represented by radial dots and line meshes, respectively. An isosurface value of 0.02 is used except for the fourth orbital from the top, where a value of 0.04 is used. The NNED and ∣∣ΔP∣∣F values are 0.083 and 2.240, respectively. Applying T1-equivalent NNED thresholds to metal diatomics and main group species reveals that strong MR character can diminish under KS-CC (Table 1). This indicates that strong correlations in MR species, including TMCs, can be captured by an alternative single reference density as long as it differs qualitatively from the canonical HF solution66. While no simple metric can guarantee improved thermochemical or electronic properties, the density deformation between HF and KS-DFT provides a useful descriptor for identifying functionals that may enhance single-reference CC methods for MR systems. However, when a given DFA yields a much more single reference picture at the CC level (measured by a sufficiently large NNED), it is safe to presume that the changes imparted by the KS density may bring improvements over HF-CC, as the KS density will be a more amenable starting point to converge the single reference CC equations. At the very least, the choice of KS-CC will not have a deleterious effect on the results if this is indeed the case. Table 1Designationsa of MR character based on HF-CC T1 diagnotics and T1-equivalent KS-based NNED metricsT1NNEDHFLDAB3LYPrevPBESCANSCAN0IBN✓✓✓✓✓✓C2✓✓✓✓✓✗BO+✓✓✓✓✓✗CN+✓✓✓✓✓✗IIbCH2✓✓✗✗✗✗O2✓✓c✗✗✗✗NF✓✓✗✓✗✗NH✗✓✗✗✗✗HFLDAPW91r2SCANPBEPBE0IIIbMn2✗✗✓✓✗✗NiO✓✓✓c✗✗✓CoCl✓✓✓✗✓✗VH✓✗✗✗✗✗Two classes of main group singlets (I & II) and one class of assorted metal diatomics (III) are examined. General NNED classifier limits are trained on TinySpins25 and BDE20.a"Check" and “cross” markers denote MR and not MR, respectively. bWhen applicable, broken-symmetry references are used. cAt the general T1-equivalent NNED threshold. Designationsa of MR character based on HF-CC T1 diagnotics and T1-equivalent KS-based NNED metrics Two classes of main group singlets (I & II) and one class of assorted metal diatomics (III) are examined. General NNED classifier limits are trained on TinySpins25 and BDE20.a"Check" and “cross” markers denote MR and not MR, respectively. bWhen applicable, broken-symmetry references are used. cAt the general T1-equivalent NNED threshold. Discussion This study clarifies how KS-DFT references can improve correlated many-body calculations relative to HF references. We demonstrate significant BDE improvements over HF-CC for a variety of TMCs with notable MR character. In particular, employing KS-DFT references with UCCSD(T) alone drastically improves the Cr2 ground-state PES topography. We also show improved spin-state energetics for bona fide MR main-group species. The concept of MR character is sometimes nebulous106 because this facet of strong correlation diminishes with more optimal reference densities. KS-DFT incorporates some correlation and resists symmetry breaking120–126 more than HF, so we posit that enhancements from KS-CC may stem from improved descriptions of electronic pair densities122. Crucially, the on-top pair density relates directly to the total electronic energy. In spin-polarized KS-DFT, broken-symmetry local spin-density approximation (LSDA) and GGA solutions yield realistic on-top pair densities and total energies127,128. For antiferromagnetic systems like Cr2, LSDA and GGA energies remain viable at the SCF level despite the symmetry dilemma from lifting variational constraints on S2. Furthermore, the reduction of self-interaction error (SIE)129–132 in a density functional is often met by a deterioration in the quality of the density and computed properties for multireference systems; such is the “zero-sum game” of DFT133. We believe these artifacts permeate the CC amplitude expressions through F~ov terms primed by PKS, as reflected in our Cr2 PEC results. SIE always affects DFT accuracy134, but the semi-canonicalization step considerably suppresses functional-driven SIE because KS and HF eigenquantities become similar. However, we find that the density and its square, proportional to the on-top density of a single determinant, must be sufficiently accurate for KS-CC to improve upon HF-CC, complicating DFA selection beyond SIE considerations. To simplify this choice, we also introduce a new difference-density metric, NNED, that addresses deficiencies in conventional MR diagnostics like T1diag and directly gauges the viability of KS densities. NNED captures features of PKS that precondition CC and MBPT methods, so NNED should guide practitioners toward DFAs that may evolve the correlated reference closer to the exact solution. NNED can be evaluated at mean-field cost while revealing a priori that certain DFAs can alleviate MR character by preconditioning the CC problem with a single reference density. For practitioners, we offer the following guidance: GGA references (PBE, PW91) consistently achieve chemical accuracy for metal–metal bonds, and we recommend them as default choices for systems with suspected MR character. For metal–ligand bonds (M-O, M-Cl, M-H, and M-H+), the choice of reference matters less, though hybrid functionals may offer marginal improvements for metal oxides. Closed-shell or d10 systems (e.g., Cu2, Zn2) show minimal sensitivity to reference choice, consistent with their negligible MR character. In general, single-reference KS-CC can handle cases of moderate MR character that challenge HF-CC. The variability in KS-CC results reflects a complicated interplay among non-Brillouin terms135–140, the transformed integrals ⟨pq∣∣rs⟩~, and the semi-canonical orbital-energy denominators. However, the epicenter of all these changes is the KS density matrix and not the KS orbitals, as it is just as valid to use localized KS orbitals or semi-canonical ones while the density remains invariant under such transformations. Examining the difference density for qualitative changes in reference determinants also underscores the importance of information-theoretic concepts and precise density reconstruction in quantum chemistry. We have clarified the physical origin of changes that KS densities induce in many-body methods. While our examples focus on single-reference KS-CC, these changes allow Thouless rotations that drive KS-CC toward the exact solution, yielding better results at lower excitation order. Specifically, we have shown that KS-CCSD(T) achieves near chemical accuracy for metal–metal BDEs at a fraction of the cost. Notably, GGA-referenced CCSD(T) produces no errors exceeding 0.51 eV across 49 transition metal diatomics, demonstrating robust transferability across diverse bonding environments (Supplementary Fig. 2). Current protocols for generating coupled cluster training data typically exclude systems that T1diag flags as multireference, introducing systematic bias toward single-reference chemistry. The NNED diagnostic offers a path to recover these systems: when NNED indicates sufficient density difference, KS-CCSD(T) can provide benchmark-quality thermochemistry data without recourse to expensive multireference methods, thereby enabling unbiased training sets that span the full spectrum of electronic structure. Finally, this work carries crucial implications for solid-state materials simulations, where HF-CCSD(T) is often dismissed due to poor performance relative to DFT. We anticipate that the NNED metric will guide DFA selection in future KS-CCSD(T) applications to solids. Methods Electronic Structure Methods We computed metal diatomic BDEs at CCSD(T)/CBS using ORCA 6.0141–145. Geometry optimizations employed SVWN5146, PBE42,43, PW91147,148, r2SCAN149,150, PBE0151, B3LYP152, ωB97X-V153, and ωB97M-V154 functionals for all dimers. For first-row transition metal systems, we used the exact-two-component (X2C) scalar relativistic approximation155,156 with the x2c-TZVPall157 basis set. We computed CCSD(T) single-point energies at experimental equilibrium distances with HF, SVWN5, PBE, PW91, r2SCAN, and PBE0 reference densities. We obtained CCSD(T)/CBS energies via two-point extrapolation of SCF158,159 and correlation energies160,161 using def2-TZVPP162 and def2-QZVPP163 basis sets. For all dimers except Mn2, we used the X2C relativistic approximation with the segmented all-electron relativistically contracted (SARC) basis sets for Coulomb fitting164, decontracted auxiliary sets for Coulomb and exchange fitting, and x2c-TZVPall for correlation fitting165. For Mn2, we followed the same procedure but with the ZORA166,167 scalar relativistic approximation and the def2-TZVP basis set for correlation fitting. For Mn, we performed two single-point energy calculations with CC-PBE0 using the ZORA and X2C relativistic approximations and the same basis sets and procedures described above. The ZORA-based energy was used for calculating the BDE of Mn2, while the X2C-based energy was used for Mn-H, Mn-H+, Mn-Cl, and Mn-O. All calculations used tight SCF convergence and the finite Gaussian-nucleus model for relativistic treatments 168. We also performed potential energy surface (PES) scans for Cr2 using the same CCSD(T)/CBS and KS-DFT protocols, except under spin polarization. We evaluated KS-DFT PES data with def2-QZVPP and compared them directly to CCSD(T)/CBS results. The Cr2 PES contained 84–124 points and used the same convergence criteria and auxiliary basis sets as in the CCSD(T) and KS-DFT calculations. We define the dissociation energy for each dimer (De) as where EM⋅ and EL⋅ are the ground-state atomic energies, EM-L is the ground-state dimer energy, and ΔESO is the spin–orbit (SO) correction. ΔESO can be obtained from experiment41,86 or first-principles calculations73, where with EM⋅,SO and EL⋅,SO the atomic SO correction energies and EM-L,SO the dimer SO correction energy. Note that when calculating the De of M-H and M-H+, we took the H energy from the two-point extrapolated UHF energy using the procedures described above. We treated several dimers with spin-polarized KS-DFT and CCSD(T)/CBS. These include Cr2 and Mn2, which have singlet antiferromagnetic ground states, and Ni2, which exhibits mixed singlet-triplet character; spin polarization improves the accuracy of the BDEs for these singlet states. We also calculated orbital energy changes with varying HFX, singlet-triplet gaps, excitation energies, NNED metrics, T1diag, and Frobenius norms of F~ov/ΔP using a development version of Q-Chem v6.2169. EPT calculations used a modified version of UQUANTCHEM170. We visualized DDNOs with IQmol171. The Supplementary Information provides additional computational details, including reference data, relativistic effects, zero-point energy corrections, and molecular geometries. Difference density analysis We formulate a density-difference metric172–174 to better assess MR character in non-canonical CC references. We define the difference between the one-electron density matrices from KS-DFT (PKS) and HF (PHF) as an approximate density cumulant: Analogous to Löwdin’s natural orbital decomposition, we solve: This yields symmetrically orthogonalized corresponding orbitals175,176 and eigenvalues representing electron density shifts in the diagonal matrix δ. The sum over positive eigenvalues, normalized to the total electron count, yields NNED, our criterion for gauging reference density similarity and a priori MR character in KS-CC. An example of NNED convergence with basis set cardinality is given in Supplementary Fig. 53. We performed this analysis using a development version of Q-Chem v6.2 169. Supplementary information Supplementary Information Transparent Peer Review file Acknowledgements A.Y.Z and K.C-F wish to thank T. Daniel Crawford for enlightening discussions on CC theory. Author contributions J.A.K. and K.C.F. conceptualized the study. B.Z., A.M.R., and A.Y.Z. collected numerical data to support the study. A.Y.Z. and K.C.F. performed the theoretical analysis to support the study. A.Y.Z. and B.Z. wrote the manuscript and all authors provided input on revisions. Peer review Peer review information : Nature Communications thanks Hyungjun Kim and the other anonymous reviewer for their contribution to the peer review of this work. A peer review file is available. Funding B.Z. acknowledges support from the National Science Foundation Graduate Research Fellowship. J.A.K. acknowledges support from the U.S. Naval Research Lab (N00173-25-1-0040). A.M.R. acknowledges support from the Wass Undergraduate Research Fellowship. This research was supported in part by the University of Pittsburgh and the University of Pittsburgh Center for Research Computing and Data through the resources provided. Specifically, this work used the H2P cluster, which is supported by NSF award number OAC-2117681. Data availability All ORCA output files, including KS-DFT optimized geometries, CCSD(T)/CBS single-point energies, atomic energies, and Cr2 potential energy surface data, are freely available on Zenodo at https://doi.org/10.5281/zenodo.17958091. Source data are provided within the linked repositories. Code availability Jupyter notebooks for parsing ORCA outputs and generating figures are available on GitHub at https://github.com/BLZ11/cc-dft. All code used for data analysis is archived at https://doi.org/10.5281/zenodo.21521932. The NNED analysis was performed using a development version of Q-Chem v6.2; the relevant routines will be included in a future public release. In the interest of broader usability, we have published a NNED analysis plugin for the open-source PySCF177 software package at https://github.com/kdcf/NNED_Analysis. Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Contributor Information John A. Keith, Email: jakeith@pitt.edu Kay Carter-Fenk, Email: kay.carter-fenk@pitt.edu. Supplementary information The online version contains supplementary material available at https://doi.org/10.1038/s41467-026-77540-x. References Associated Data Supplementary Materials Supplementary Information Transparent Peer Review file Data Availability Statement All ORCA output files, including KS-DFT optimized geometries, CCSD(T)/CBS single-point energies, atomic energies, and Cr2 potential energy surface data, are freely available on Zenodo at https://doi.org/10.5281/zenodo.17958091. Source data are provided within the linked repositories. Jupyter notebooks for parsing ORCA outputs and generating figures are available on GitHub at https://github.com/BLZ11/cc-dft. All code used for data analysis is archived at https://doi.org/10.5281/zenodo.21521932. The NNED analysis was performed using a development version of Q-Chem v6.2; the relevant routines will be included in a future public release. In the interest of broader usability, we have published a NNED analysis plugin for the open-source PySCF177 software package at https://github.com/kdcf/NNED_Analysis.