WorksPhysics and Astronomy
Brueckner Orbitals in Local Pair Natural Orbital Coupled-Cluster Theory: DLPNO-BCCD(T) and DLPNO-BCCD(T)_Lambda2
Abstract
In this work, we implement closed-shell Brueckner orbitals for the efficient domain-based local pair natural orbital DLPNO-CCSD(T) and DLPNO-CCSD(T) methods of the free and open-source Psi4 package. We introduce a new algorithm that allows Brueckner orbitals to be computed within the DLPNO framework, resulting in the novel DLPNO-BCCD(T) and DLPNO-BCCD(T) methods. We show that DLPNO-BCCD(T) approximates its canonical BCCD(T) counterpart with significantly lower errors relative to the errors incurred by DLPNO-CCSD(T) over CCSD(T). Through a test on the S22 set of noncovalent dimers, the peak error across the set is reduced from 0.36 kcal mol to 0.25 kcal mol using normal cutoffs, and 0.28 kcal mol to 0.13 kcal mol at tight cutoffs. In addition, we show that BCCD(T), a -only implementation of Brueckner orbitals for the asymmetric triples method, generally produces the best agreement with CCSDT, compared to the (T), (T), and B(T) perturbative triples corrections. This is established through benchmarks on thermochemistry tests on the barrier heights of semibullvalene and bullvalene Cope rearrangement, while the robustness of DLPNO-BCCD(T) is demonstrated along bond-breaking potential energy curves, where it matches the orbital-optimized quasi-variational asymmetric renormalized-triples OQVCCDAR(T) method closely for single-bond breaking, and approximates DLPNO-CCSDT respectably even when triple bonds are stretched, at a fraction of the computational cost.
Introduction
Much of the predictive power in quantum chemistry has been attributed to the development of coupled-cluster (CC) theory, [1, 2] with accurate predictions for molecular energetics and properties often within experimental errors. This is because CC theory provides a size-extensive—and for systems dominated by dynamic correlation, rapidly convergent— approach to approximate the full configuration interaction (FCI) wavefunction [3, 4, 5] through truncation at a given level of electronic excitation. CC theory uses the exponential ansatz:
with the energy given by
where represents a single reference Slater determinant, typically derived from Hartree–Fock, but alternative references, such as determinants from density functional theory, are also resurgent. [6, 7, 8] The cluster operator represents the sum of all electronic excitations considered from the reference determinant, up to a given excitation level, . The size-extensivity of coupled-cluster theory arises from this ansatz, since the exponential operator allows “disconnected” products of lower-level excitations to be a proxy for higher-level excitations, a property not present in truncated configuration interaction (CI) approaches. For example, in CC with single and double excitations (CCSD), [9] where , much of the effect of CI quadruples excitations is encapsulated in the , , and terms that arise from , providing both size-extensivity and improved convergence towards the FCI wavefunction.
In the CC hierarchy, considering up to triple excitations (), as done by the CCSDT method, [10, 11, 12, 13] is sufficient to capture almost all of the chemically meaningful correlation effects in typical systems. Comparisons to FCI [11, 14, 15] and high-order CC [16, 17, 18, 19] results show that even for very demanding quantities such as atomization energies, post-CCSDT contributions for small molecules are typically below 1 kcal mol, the customary threshold of “chemical accuracy”. The errors of CCSDT are generally smaller still for barrier heights [19, 20] and non-covalent interaction energies. [19, 21, 22, 23, 24]
Unfortunately, this exact consideration of triples is prohibitively expensive, having an iterative cost with respect to system size . This led to the development of the renowned “gold standard” CCSD(T) method, [25, 26, 27, 28] where single and double excitations are treated with CCSD ( and ), at iterative cost, while the triples contribution () is estimated using perturbation theory at a non-iterative cost. In cases where the Hartree–Fock reference is a reasonably good approximation of the FCI wavefunction of the molecule, CCSD(T) gives results very close to CCSDT, and even performs better than CCSDT when compared against CCSDTQ due to (T) typically overestimating the triples contribution, which largely cancels the neglected connected quadruples. [24, 29, 30, 31, 32, 33] This is further reflected in the good performance of CCSD(T) and CCSD(T)-based composite schemes for thermochemistry, [17, 30, 34, 35, 36, 37, 38] equilibrium geometries and spectroscopic constants, [32, 39, 40, 41, 42, 43, 44, 45, 46] as well as magnetic properties. [47, 48, 49] CCSD(T) interaction energies of small non-covalent complexes also agree with quantum Monte Carlo (QMC) results within about 0.1–0.2 kcal mol. [50, 51, 52]
However, in multi-reference cases such as molecules with stretched geometries, the Hartree–Fock orbitals are no longer a reliable reference, and they lead to the (T) correction producing unphysical values. [15, 25, 27, 33, 53] The inclusion of full triples and higher-order excitations can ameliorate many of the shortcomings of a bad Hartree–Fock reference, [54, 55, 56, 57, 58, 59] as can, to a good extent, improving the perturbation theory estimate through left-hand amplitudes in the CCSD(T) approach, [33, 60, 61, 62, 63, 64, 65, 66] whose defining step is solving for these amplitudes by finding the stationary point of the Lagrangian
with respect to the coupled-cluster amplitudes. By computing the left-hand eigenvector of the similarity-transformed Hamiltonian , CCSD(T) avoids the assumption inherent in CCSD(T) that is an accurate approximation, which deteriorates when the non-Hermiticity of escalates in multireference cases. [61, 67, 68]
An alternative to the conventional Hartree–Fock reference () is to incorporate electron-correlation effects into the determination of the molecular orbital (MO) coefficients. One approach is orbital optimization, as in orbital-optimized CCD (OO-CCD) [69, 70, 71] and orbital-optimized quasi-variational CCD (OQVCCD) [72], where the orbitals are optimized to make the total CC energy functional stationary with respect to orbital rotations. A different approach can be derived by noting that while a non-truncated CC (or CI) expansion yields the exact correlated wavefunction from any reference determinant satisfying , some references have a greater overlap with than others.
It can be shown through Thouless’s theorem [73] that the CI coefficients of single excitations are nil for any reference which maximizes . Since for an non-truncated CC expansion in intermediate normalization, the leading order coefficients are the amplitudes, is a necessary (albeit not sufficient) condition for the reference determinant to have maximal overlap with . This condition we consider as the exact Brueckner condition, while the determinants that satisfy it are the exact Brueckner determinants . [74, 75] Equivalently, the overlap of with is stationary with respect to occupied–virtual orbital rotations.
Truncating the CC expansion leads to approximate Brueckner conditions, where yields approximate Brueckner determinants that no longer have maximal or stationary overlap with . This of course does not preclude approximate Brueckner orbitals from being an improvement over a Hartree–Fock reference. The simplest approximate Brueckner method is the Brueckner coupled-cluster doubles (BCCD) approach. [76, 77] Here, an approximate is determined self-consistently by repeatedly rotating the orbitals using amplitudes from CCSD iterations until the approximate Brueckner condition is achieved. At convergence, the BCCD energy is
With the improved reference, the foundation of perturbative triples is improved as well, with the BCCD(T) method [78, 79, 80] showing improved resilience against non-variational behavior at stretched bond lengths, [27, 62, 81] and the closely related OO-CCD(T) reinforces that reference relaxation can rescue (T). [65] This improved performance at stretched bond lengths suggests that even obtained under an approximate Brueckner condition may represent a better zeroth-order wavefunction than .
In this research, we investigate , [62, 82] which we believe is a further improvement upon both the BCCD(T) and approaches for a perturbative triples algorithm. The central contribution of this work is the implementation of the method, which introduces a novel algorithm that enforces the Brueckner condition in the local pair natural orbital space while maintaining the locality of the occupied molecular orbital space.
For our algorithm, we consider the BCCD approach, using the interpretation of the method using response amplitudes computed at the CCD level of theory computed using a Brueckner reference optimized with CCSD. This means we do not explicitly consider the effects of . Taube has shown that when CCSD is computed with a Brueckner reference, does not imply when computing response amplitudes. [82] When is also computed with a BCCD reference, the approach is referred to as BCCSD by Taube. In this work, we only consider the DLPNO-BCCD(T), or as we shall name it, the DLPNO-BCCD(T) algorithm. A future version of our method that also considers will be known as DLPNO-BCCD(T). The equations are also solved in the local pair natural orbital basis following our prior work. [83]
In this work, we show that BCCD(T) produces a perturbative triples estimate that has much higher fidelity to CCSDT benchmarks compared to CCSD(T), and BCCD(T) through tests on a variety of bond-breaking curves involving single, double, and triple bonds. For these computations, DLPNO-BCCD(T) can effectively be treated as canonical due to the tight pair natural orbital cutoffs used (). In addition, we show that DLPNO-BCCD(T) shows much better agreement with best canonical CCSDT estimates compared to DLPNO-CCSD(T), DLPNO-BCCD(T), and DLPNO-CCSD(T) for Cope Rearrangement of bullvalene/semibullvalene. Furthermore, we investigated the magnitude of of local correlation errors for DLPNO-BCCD(T) compared to canonical BCCD(T), and find that at a given set of cutoffs, DLPNO-BCCD(T) produces much lower errors compared to DLPNO-CCSD(T) through tests on non-covalent interaction (NCI) energies in the S22 set. [84] To demonstrate the large-system capabilities of our code, we also ran computations on a coronene dimer and an octadecane dimer from the L7 NCI set. [85]
Notation
We use the following conventions to describe the indices of matrices and tensors appearing in this work, analogous to the notation used in prior research: [86, 87]
, , , : atomic orbitals; these range from 1 to , the number of basis functions
, , , : canonical and local occupied molecular orbitals; these range from 1 to , the number of occupied orbitals
, , , : canonical virtual molecular orbitals; these range from 1 to , the number of virtual orbitals
, , , : general canonical molecular orbitals; these range from 1 to
, , , : projected atomic orbitals; these range from 1 to
, , , : projected atomic orbitals localized to pair ; these range from 1 to , number of PAOs local to LMO pair
, , , : projected atomic orbitals localized to triple ; these range from 1 to , number of PAOs local to LMO triple
, , , : “Interacting” local molecular orbitals in a pair domain , defined as the set of all occupied orbitals such that and forms a valid pair; these range from 1 to , number of LMOs in the domain of LMO pair
, , , : Pair natural orbitals in each pair domain ; these range from 1 to , number of PNOs in the domain of LMO pair
, , : “Interacting” local molecular orbitals in a triples domain , defined as the set of all occupied orbitals such that , , and are all strong or weak pairs; these range from 1 to , number of LMOs in the domain of LMO triplet
, , , : Triples natural orbitals in each triples domain ; these range from 1 to , number of TNOs in the domain of LMO triplet
, : auxiliary basis functions for density-fitted ERIs; these range from 1 to , number of auxiliary basis functions for density fitting
, : local auxiliary basis functions in each pair domain ; these range from 1 to , number of auxiliary basis functions local to LMO pair
, : local auxiliary basis functions in each triples domain ; these range from 1 to , number of auxiliary basis functions local to LMO triplet
The relative sizes of these indices are typically:
where is the system size represented by the number of atoms or electrons.
Theory
Brueckner Orbital Construction in Local Pair Natural Orbital Basis
For implementing Brueckner orbitals in DLPNO coupled-cluster theory, we iteratively rotate the reference orbitals using information from successive DLPNO-CCSD computations, [86] until the norm of the -amplitudes drops under a given tolerance (typically –). In our implementation, our choice of norm is the maximum absolute value. For DLPNO in particular, there is a caveat: after the Brueckner orbitals are computed, the occupied space needs to be re-localized each iteration. A summary of the major steps of this algorithm is depicted in Figure 1. Our approach can be denoted as the “Local Brueckner Optimization Algorithm”, and the orbitals obtained with this approach as “local Brueckner orbitals”, to distinguish them from orbitals obtained from canonical CCSD amplitudes.

Figure 1. Simplified flowchart of a typical computation with the DLPNO-Brueckner orbital optimization algorithm. Generative AI assistance (GPT-5.6 Sol and Claude 5 Opus) was used to design the flowchart template, and the authors manually edited the template and verified scientific accuracy.
To start, a series of standard DLPNO-CCSD computations are performed, with the initial macro-iterations limited to typically 5–10 CCSD micro-iterations until the norm of the amplitudes drops below , after which case the CCSD iterations are run until convergence– in each Brueckner macroiteration. This is done to improve numerical stability in problems with unstable or large amplitudes in the initial CCSD macro-iterations. Each time a DLPNO-CCSD computation is completed, the amplitudes of each active occupied molecular orbital , computed in the diagonal PNO space of pair (defined as ), are projected to the global canonicalized PAO space, mathematically equivalent to the virtual space from the corresponding Hartree-Fock computation. This projection can be defined through:
Next, the anti-symmetric rotation generator matrix is formed as a block matrix between the occupied and virtual spaces:
Subsequently, an exponentiation of this matrix is performed, with the intermediate .
Once this matrix exponentiation is performed, a new set of occupied orbital coefficients can be obtained from the blocks of this matrix. Note that denotes a successive iteration, and denotes the preceding iteration:
where represents an element from the occupied-occupied, or block of , and comes from the resulting occupied-virtual () block. represents the occupied orbital coefficients from Brueckner iteration (localized molecular orbitals), while represents the corresponding virtual orbital coefficients, which are computed from canonicalizing the global projected atomic orbital (PAO) space. Once these new occupied orbital coefficients are recomputed, they need to be renormalized:
After these new coefficients are recomputed, it is important to recompute the Fock matrix, as it changes after the orbitals are updated. In the RHF case,
where refers to the atomic orbital (AO) density matrix computed from the updated orbital coefficients in Brueckner iteration
The Hartree-Fock energy also needs to be recomputed, which in the RHF case, is
Finally, the new occupied molecular orbital coefficients () are recanonicalized in order to form a diagonal representation of the Fock matrix in the occupied space. After this, they are re-localized– using the same algorithm selected for the initial DLPNO-CCSD computation– to form a new set for localized molecular orbitals (LMOs). Finally, the DLPNO-CCSD computation is rerun with the new set of LMOs.
The new corresponding virtual space (in PAOs), formed from subtracting out the span of the new LMOs (denoted by C′) from the global atomic orbital space, and then normalized (Equation 17).
The computation then proceeds through the remaining steps of a DLPNO-CCSD computation in Psi4, which includes the dipole, semi-canonical MP2, and PNO-LMP2 pre-screening steps–as well as the formation of pair natural orbitals (PNOs)–then solving the DLPNO-CCSD equations in the new basis. A more detailed overview of the DLPNO-CCSD algorithm in Psi4 is provided in Reference 86. For this algorithm, every “weak pair” that normally arises from the original algorithm in Ref. 86 will be treated as a strong pair. For consistency, all DLPNO-CCSD(T), DLPNO-CCSD(T), DLPNO-BCCD(T), and DLPNO-BCCD(T) results presented in this work will use this “no weak pairs” treatment.
Typically, the PNO ranks increase by 1-2 PNOs per pair, on average, between the initial DLPNO-CCSD computation and the final DLPNO-BCCD result. This computation is considered converged when the norm of the amplitudes () falls below a given Brueckner tolerance, defined to be (). A default of is sufficient for most applications, though we used a of for most of the computations used in this work, with used for all the bond breaking curves used in this work. In this work, we went with , or the max absolute value. In addition, the local Brueckner orbital optimization usually converges within around 5-8 DLPNO-CCSD macroiterations, independent of the size of the systems studied.
Non-Brillouin Modifications
Brueckner orbitals do not obey/satisfy the Brillouin condition: . Thus, in this formalism, we need to introduce a new matrix, in the LMO/PAO basis (and update the DLPNO-CCSD working equations accordingly):
MP2 Terms
The MP2 expression for pre-screening now includes an additional term from :
This singles Fock matrix contribution contributes to the energy of the semi-canonical MP2 “screened pairs” as well as the PNO truncation correction computed to compensate for the PNO truncation error for diagonal pairs .
CCSD Terms
Certain coupled-cluster equations will look different here compared to our original DLPNO-CCSD algorithm, of reference 86, with the inclusion of the non-Brillouin terms. Equations 99 and 100 now become (from our original work, in reference 86):
The CC energy expression also receives an update:
Triples Terms
In addition, certain triples intermediates need to be augmented with contributions from these terms. For DLPNO-BCCD(T), the intermediate from DLPNO-CCSD(T) (Equation 110 in Ref. 86) is augmented with:
For DLPNO-BCCD(T), the is augmented accordingly (Equations 96 and 97 in Ref. 83).
In these equations represents the “short permutation operator”, defined through P_{S}X_{ijk}^{abc}=X_{ijk}^{abc}+X_{jik}^{bac}+X_{kji}^{cba}.
Implementation Details and AI Disclosure
The initial working version of DLPNO-BCCD(T) was implemented without AI assistance, and the results on the bond-breaking reactions (§ 8.A) used this initial implementation https://github.com/psi4/psi4/commit/46822f951c9bb6bf625537ac3456bacebd4ac9ec.
For the other benchmarks on the larger computations, GPT-5.6 Sol was utilized for implementing features that improved the robustness of the local Brueckner code, such as DIIS improvements, [88] as well as implementing a “frame transport” scheme for localizing orbitals between Brueckner macroiterations involving a Procrustes solution, [89] and a Hungarian maximum overlap assignment-based scheme [90] to preserve orbital permutations and phases. In addition, augmented Hessian [91, 92] variants of the localization algorithms (Foster–Boys, Pipek–Mezey, and IBO) [93, 94, 95, 96] were implemented. These latest code changes are incorporated in https://github.com/andyj10224/psi4/tree/brueckner_lambda. All other results (§ 8.B, V.C, and V.D) used this updated version of the code given by the branch above. A GPT-5.6 Sol Max audit also corrected a DIIS bug in the left-hand coupled-cluster solver from our prior work on DLPNO-CCSD(T), [83] where the provided error vector was the amplitudes themselves and not the corresponding residual. Though this does not affect numerical results, this fix drastically improves convergence of the equations. For our tests, with the bond-breaking scenarios, a Brueckner orbital convergence tolerance of was used for the amplitudes, and for all others, is used.
Generative AI (GPT 5.6 Sol and Claude 5 Opus) was also used in preparing Figure 1 with the providing of a flowchart template through an .svg file, through which the final figure was rendered by the authors. In addition, GPT 6 Astra was used in creating the Python Numpy/Matplotlib source code that rendered Figure 8. In addition, GPT 5.6 was used for a literature search on the DIIS improvements, frame transport, and augmentedHessian implementations of orbital localization algorithms as well as their implementation (prompts provided in the supporting information). GPT 6 Astra was also used during part of the editing procedure for many of the captions in the supporting information, though no scientific content was modified.
Results
Bond-Breaking Scenarios
To test the correctness and capabilities of our DLPNO-BCCD(T) and DLPNO-BCCD(T) theories and prototype implementations, we apply them in a battery of challenging bond-breaking scenarios:
Breaking a C–H or the C–Cl bond of CHCl, with a F spectator at 200 Å
Breaking the bonds of N, F, S, and HF
Breaking the triple bond of CH
We also compare our new methods to the Molpro [97] implementations of canonical CCSD(T) and BCCD(T), as well as the quasi-variational coupled cluster method with canonical triples (OQVCCD(T)), [81, 98] renormalized triples (OQVCCDR(T)), [99] and asymmetric renormalized triples (OQVCCDAR(T)). [100, 101] Furthermore, we anchor our comparisons to higher-order coupled cluster theory by computing CCSDT results using Psi4’s prototype DLPNO-CCSDT implementation. [87]
In all cases we use the double- basis sets of Dunning, augmented with diffuse functions on all atoms, and with tight d-functions on second row atoms. We minimize the impact of the density fitting approximations employed in our DLPNO methods by using fitting bases one higher: aug-cc-pVTZ-JKFIT and aug-cc-pVTZ-RI. To negate the effects of differing total energies, all comparisons use energies relative to the equilibrium energies of the systems, computed with the method in question. The precise computational details and full output files are provided in the Supporting Information.
CHCl bond breaking with F spectator
A major motivation for developing DLPNO-BCCD(T) is the need for an affordable and black-box quantum chemistry backend for the automated development of full-dimensional reactive potential energy surfaces (PES) using active learning systems such as Robosurfer. [102] In this application, the user of the quantum chemistry program is not a person, but another program that requests energy evaluations wherever the current training set of the fitted PES appears to be deficient. In practice, this can lead to quantum chemistry computations at arbitrary geometries. If the method in use exhibits severe pathologies (e.g. non-variational divergence or notable discontinuities in energy) at extreme geometries, such data can easily derail the active learning process and lead to the failure of PES development. Therefore, we chose the C–H or the C–Cl bond breaking potential energy curves (PECs) in the reactant asymptote of the CHCl + F reaction as our first test case, with the aug-cc-pV(D+d)Z basis set. Beyond its direct relevance to PES development applications, these unusual geometries should also exercise the DLPNO implementation, as the F ought to have negligible interactions with the CHCl where the bond-breaking processes take place.
Validation of the correspondence between Psi4 and Molpro energies
Given our intention of comparing the results from two different quantum chemistry packages, we hold it prudent to begin by validating that all differences seen are attributable to the differences in CC theories and not implementation details of integral computation or similar factors. We computed the two PECs with canonical MP2 and BCCD(T) both in Psi4 and Molpro (Fig. S1-S4), and found that the difference between the MP2 relative energies is < kcal mol for C-H distances in Å and for C-Cl distances in Å. The canonical BCCD iterations in Psi4 did not converge at C-H distances > Å and at C-Cl distances > Å, but no noteworthy discrepancies are observed in the Å and Å ranges, respectively. Therefore, we conclude that comparing our new DLPNO-BCC results to Molpro reference values generally rests on solid foundations.
C-H and C-Cl bond breaking
The PECs in Figure 2 show qualitatively similar behavior between both bonds for all methods, with the notable exception of CCSD(T), which exhibits a stark non-variational collapse when the C-Cl bond is stretched beyond , arguably performing worse than MP2 in this particular case. All other CC methods yield at least somewhat reasonable PECs, with CCSD(T) remaining the least qualitatively correct in the C-H case, exhibiting the largest unphysical local maximum. Methods eschewing triple excitations (CCSD, BCCD) produce qualitatively correct dissociation curves with no local maxima, but at the cost of substantially overestimated dissociation energies and inferior accuracy at geometries near equilibrium.
Figure 2. Potential energy curves of C-H (upper panel) and C-Cl (lower panel) bond breaking in CHCl bond breaking with a F spectator at a distance of , computed with Psi4 and Molpro using MP2, CCSD, CCSD(T), OQVCCD(T), OQVCCDAR(T), DLPNO-BCCD(T), DLPNO-BCCD(T) and DLPNO-CCSDT, with the aug-cc-pV(D+d)Z basis set. All energies except the DLPNO and OQVCCDAR(T) results are relative to the energies computed by the given quantum chemistry package at the equilibrium geometry optimized by Molpro 2023 using the same level of theory. DLPNO and OQVCCDAR(T) PECs use BCCD(T) geometries.
The effect of using Brueckner orbitals is modest at the doubles level, for the C-H PECs and for C-Cl (Fig. S5, S6), but is greatly amplified if perturbative triples are used, growing to along the C-H PEC and already exceeding at a C-Cl distance of . Indeed in both cases, BCCD(T) yields dissociation curves with only a moderate unphysical hump and cures the nonvariational collapse of CCSD(T) at long C-Cl distances.
A unique alternative to ordinary CC methods is the family of quasi-variational CC methods implemented in Molpro. Briefly, these modify the doubles equations using terms derived from variational coupled cluster theory, improving their robustness while maintaining two-electron/two-hole exactness, size extensivity, and other desirable theoretical qualities. The effects of singles are included via orbital optimization, yielding the orbital-optimized quasi-variational doubles (OQVCCD) as a base method. This may then be followed by the ordinary perturbative triples (OQVCCD(T)), a renormalized asymmetric perturbative triples (OQVCCDAR(T)), or a symmetric approximation of the latter (OQVCCDR(T)).
OQVCCD(T) does not offer a meaningful improvement over BCCD(T), matching its relative energies within for C-H bond breaking and for C-Cl (Fig. S5, S7). However, its renormalized variants perform extremely well, producing qualitatively almost correct dissociation curves with a hump height of only . This has a reasonably good quantitative agreement with the DLPNO-CCSDT dissociation energy, a remarkable result from a method without iterative triples, suggesting that both orbital relaxation and an improvement over the canonical (T) formula are required for robust single bond breaking. The symmetric approximation of the renormalized triples has a negligible effect (Fig. S8, S9) on the energies along these PECs.
Our new DLPNO-BCCD(T) implementation matches the relative energies of canonical BCCD(T) within (Fig. S8, S9), supporting the correctness of our implementation, and demonstrating that at least with the set of thresholds used here, the DLPNO approximation error is negligible in BCCD(T). We have also tested the effect of the semicanonical triples terms by rerunning the Molpro BCCD(T) computations without them, and found peak differences of 0.19 and 0.36 kcal mol over the C-H and C-Cl PECs, respectively, confirming the necessity of implementing these terms in our DLPNO-BCCD(T) code.
performs excellently, matching the OQVCCDAR(T) results almost exactly for C-H bond breaking (Fig. S5) and only deviating from it by 0.34 kcal mol at a C-Cl distance of 6.25 Å. This also results in matching DLPNO-CCSDT within 1.6 kcal mol over the C-H PEC and 2.96 over the C-Cl PEC, with the deviation shrinking again near the dissociation limit.
F dissociation
Stretching the single bond of F is superficially similar to the C-H stretch scenario discussed above for CHCl, but with some notable differences. The magnitude of CCSD(T)’s non-variational behavior is a much greater fraction of the dissociation energy, and unlike the case of C-H stretch OQVCCD(T) and BCCD(T) are not near-equivalents, as here BCCD(T) is a slight improvement over CCSD(T) while OQVCCD(T) performs even worse than CCSD(T). Substituting the triples approximation in OQVCCD(T) with one of the renormalized variants yields a major improvement and the resulting dissociation curves are almost qualitatively correct, with only a modest (~2.1 kcal mol) unphysical maximum and a good agreement with the CCSDT dissociation limit inferrable from the section of the PEC where DLPNO-CCSDT converges.
Our DLPNO-BCCD(T) results reproduce the canonical values with negligible error (< 0.035 kcal mol, Fig. S11) at bond distances ≥ 0.6Å, and performs excellently, offering PECs similar to OQVCCDAR(T) in a linear-scaling framework.
HF dissociation
The dissociation of HF is a good probe for the behavior of electronic structure theories in the limit of breaking a single bond with a large difference in electronegativities. The resulting PECs are similar to those obtained for C-H bond breaking in CHCl (Figure 4): CCSD(T) yields an unphysical local maximum, but no catastrophic non-variational breakdown, introducing orbital relaxation effects through BCCD(T) or OQVCCD(T) results in a similar and substantial improvement. Finally, replacing the ordinary (T) expression with one of the renormalized or variants produces nearly identical, qualitatively correct dissociation curves.
Figure 4. Potential energy curves of HF bond breaking, computed with Psi4 and Molpro using BCCD(T), DLPNO-BCCD(T), DLPNO-BCCD(T), CCSD(T), OQVCCD(T), OQVCCDR(T), OQVCCDAR(T) and DLPNO-CCSDT, with the aug-cc-pVDZ basis set. All energies except the DLPNO results are relative to the energy at the equilibrium geometry optimized using the same level of theory. DLPNO PECs are relative to the single-point energies at the BCCD(T) geometry.
Figure 3. Potential energy curves of F bond breaking, computed with Psi4 and Molpro using BCCD(T), DLPNO-BCCD(T), DLPNO-BCCD(T), CCSD(T), OQVCCD(T), OQVCCDR(T), OQVCCDAR(T) and DLPNO-CCSDT, with the aug-cc-pVDZ basis set. All energies except the DLPNO results are relative to the energy at the equilibrium geometry optimized using the same level of theory. DLPNO PECs are relative to the single-point energies at the BCCD(T) geometry.
Our DLPNO-BCCD(T) results match canonical BCCD(T) practically exactly, within 0.013 kcal mol, while DLPNO-BCCD(T) (together with the renormalized QVCC variants) matches the DLPNO-CCSDT PEC excellently.
Singlet dissociation
The dissociation of singlet is substantially more difficult to treat than the single bond breaking processes discussed above, and single-reference coupled cluster methods more truncated than CCSDTQ are not expected to provide an accurate PEC. Furthermore, the RHF wavefunction of singlet already exhibits spatial symmetry breaking at equilibrium bond length, obeying only the point group, instead of the full symmetry expected for diatomics, making this system an interesting test case.
As expected on theoretical grounds, the DLPNO-CCSDT dissociation curve is imperfect and exhibits a small unphysical local maximum. The non-variational behavior of CCSD(T) is much more severe, albeit not to the catastrophic extent seen for C-Cl bond breaking in CHCl. Unlike all previous systems, Brueckner orbitals by themselves offer no improvement for S, as BCCD(T) performs slightly worse than CCSD(T), while OQVCC methods yield considerably different results than BCCD(T). At an S distance of 3.1 Å near the local maximum of the BCCD(T) and DLPNO-CCSDT PECs, we could not converge the OQVCC iterations, but we note that OQVCCD(T) and its renormalized variants all show a curvature discontinuity near this point. Beyond this anomaly, all OQVCC methods produce dissociation limits considerably higher than DLPNO-CCSDT, especially the renormalized methods.
Our DLPNO-BCCD(T) method approximates the relative energies of its canonical counterpart within 0.045 kcal mol, supporting the view that all components of the DLPNO-BCC implementation remain viable. Unlike the previously discussed singly-bonded systems, S has a DLPNO-BCCD(T) PEC that does not resemble OQVCCDR(T) or any other method, but with a tolerable unphysical maximum, no discontinuity and a dissociation energy that is the closest to DLPNO-CCSDT among all tested methods, it is arguably the best-behaved method in this test.
Figure 5. Potential energy curves of S bond breaking, computed with Psi4 and Molpro using BCCD(T), DLPNO-BCCD(T), DLPNO-BCCD(T), CCSD(T), OQVCCD(T), OQVCCDR(T), OQVCCDAR(T) and DLPNO-CCSDT, with the aug-cc-pV(D+d)Z basis set. All energies except the DLPNO results are relative to the energy at the equilibrium geometry optimized using the same level of theory. DLPNO PECs are relative to the single-point energies at the BCCD(T) geometry.
N dissociation
The dissociation of N’s triple bond is a brutally challenging problem for any single-reference quantum chemistry method, the full description of which would require CCS-DTQPH, a method with bleak prospects of practicality even in the most efficient local correlation frameworks. Therefore, none of the methods developed in this work can reasonably be expected to provide good dissociation curves, but the comparative severity and nature of their pathologies should still be informative.
As expected, CCSD(T) exhibits strong non-variational breakdown above N-N distances of Å, and unlike the single bond-breaking scenarios in CHCl Brueckner orbitals offer no improvement and indeed slightly bring forward non-variational behavior. The family of quasi-variational methods performs exceptionally well considering the difficulty of this challenge. All of them produce qualitatively correct dissociation curves until they become impossible to converge at distances above 2.3 Å, outperforming even DLPNO-CCSDT, which suffers from non-variational behavior as much as BCCD(T) until ~2.05 Å and only outperforms CCSD(T) after ~2.05 Å.
Figure 6. Potential energy curves of N bond breaking, computed with Psi4 and Molpro using BCCD(T), DLPNO-BCCD(T), DLPNO-BCCD(T), CCSD(T), OQVCCD(T), OQVCCDR(T), OQVCCDAR(T) and DLPNO-CCSDT, with the aug-cc-pVDZ basis set. All energies except the DLPNO results are relative to the energy at the equilibrium geometry optimized using the same level of theory. DLPNO PECs are relative to the single-point energies at the BCCD(T) geometry.
Our DLPNO-BCCD(T) PEC matches Molpro’s canonical almost perfectly (Fig. S10) with an error < 0.075 kcal mol in relative energy over the PEC, and although DLPNO-BCCD(T) still yields an unphysical local maximum on the PEC, it outperforms CCSD(T) and DLPNO-CCSDT and produces the best PEC outside of the quasi-variational framework. Our results support the conclusions of the prior work of Taube and Bartlett on using on N dissociation. [62]
CH triple bond breaking
Stretching the triple bond of acetylene is barely any kinder than N on electronic structure theory: CCSD(T) shows strong non-variational behaviour, BCCD(T) performs slightly worse still, and while DLPNO-CCSDT is a definite improvement, it still exhibits a substantial unphysical maximum. QVCC methods fare much better, and all yield qualitatively correct dissociation curves, with the dissociation energy being much higher for the renormalized variants. The peak error in relative energy introduced by our DLPNO-BCCD(T)
Figure 7. Potential energy curves of breaking the C-C bond of , computed with Psi4 and Molpro using BCCD(T), DLPNO-BCCD(T), DLPNO-BCCD(T), CCSD(T), OQVCCD(T), OQVCCDR(T), OQVCCDAR(T) and DLPNO-CCSDT, with the aug-cc-pVDZ basis set, from the Hartree-Fock solution connected to the equilibrium Hartree-Fock solution. All energies except the DLPNO results are relative to the energy at the equilibrium geometry optimized using the same level of theory. DLPNO PECs are relative to the single-point energies at the BCCD(T) geometry.
implementation is again very small ( kcal mol, Fig. S15), supporting that, with the very tight thresholds used here, the DLPNO approximations remain valid even in the face of extreme electronic structure challenges. Considering its cost, DLPNO-BCCD(T) performs outstandingly well at dissociating , with its PEC closely following the DLPNO-CCSDT curve.
Cope rearrangement barrier heights
For our first test on larger systems (Table I), we evaluate each of our available triples methods for the Cope rearrangement barrier heights of semibullvalene (), and bullvalene (), with the cc-pVDZ(p,p) basis, which is cc-pVDZ without polarization functions for heavy atoms, and with the full cc-pVDZ basis sets. The geometries for the equilibrium and barrier heights, as well as canonical CCSDT-CCSD(T) differences for all computations except bullvalene/cc-pVDZ, are derived from the groundbreaking work of Karton [103]. At the smaller cc-pVDZ(p,p) basis, we observed that the canonical CCSDT–CCSD(T) and DLPNO-CCSDT-CCSD(T) differences are conserved across semibullvalene and bullvalene, and therefore provide our best estimate for canonical CCSDT-CCSD(T) for bullvalene/cc-pVDZ through [bullvalene/cc-pVDZ] [bullvalene/cc-pVDZ] + ( [semibullvalene/cc-pVDZ] − [semibullvalene/cc-pVDZ]). From these results, we show that DLPNO-BCCD(T) consistently produces the closest results to canonical T-(T) differences among the featured perturbative triples methods (differences of +0.07 and +0.08 kcal mol for the cc-pVDZ(p,p) computations, and −0.03 and +0.03 kcal mol for the cc-pVDZ computations). This also highlights that B(T) is indeed capturing higherorder triple effects compared to (T). in and of itself over corrects for (T), and can give worse answers relative to Full T compared to (T), as shown in our prior work [83] and here, while adding the Brueckner orbitals stabilizes this effect to provide the “sweet spot”.
| Reaction | |||||
| Semibullvalene ()/cc-pVDZ(p,p) | +0.38 | −0.18 | +0.21 | +0.12 | +0.14 |
| Bullvalene ()/cc-pVDZ(p,p) | +0.42 | −0.16 | +0.27 | +0.17 | +0.19 |
| Semibullvalene ()/cc-pVDZ | +0.40 | −0.11 | +0.30 | +0.36 | +0.33 |
| Bullvalene ()/cc-pVDZ | +0.46 | −0.07 | +0.39 | +0.39 | +0.36* |
Table I. Comparing triples energy corrections compared to (T) for barrier heights for the Cope rearrangement of semibullvalene and bullvalene, cc-pVDZ(p,p) and full cc-pVDZ basis, presented in kcal mol. , , , and represent, respectively, the difference of DLPNO-CCSD(T), [83] DLPNO-BCCD(T), DLPNO-BCCD(T), and DLPNO-CCSDT [87] results from computed DLPNO-CCSD(T) results. represents a canonical T-(T) difference computed from the work of Karton in 2020 (Ref. [103]). Note that the canonical bullvalene/cc-pVDZ result for T-(T) is not available and estimated with [bullvalene/cc-pVDZ] + ( [semibullvalene/cc-pVDZ] − [semibullvalene/cc-pVDZ]).
Non-Covalent Interactions
A good test for the accuracy of local correlation methods is non-covalent interactions, where local correlation errors can have a greater commensurate effect on relative energies. We compare our DLPNO-BCCD(T) algorithm with canonical BCCD(T) at normal (“NormalPNO”) and tight (“TightPNO”) cutoffs; and also compare DLPNO-CCSD(T) computed at the same cutoffs compared to canonical CCSD(T), on the S22 data set of non-covalent dimers, in the cc-pVDZ basis. The counterpoise (CP) correction is applied for each system. [84] Individual results are presented in Table II and summary statistics are presented in Figure 8, grouped by interaction type. Dimers S22-1 to S22-7 represent predominantly hydrogen-bonded dimers (HB), dimers S22-8 to S22-15 represents a dispersion-dominated complex (DD), while dimers S22-16 to S22-22 represent mixed influence systems (MIX).
| Interaction | ||||||
| Ammonia Dimer (S22-1) | −1.76 | +0.02 | +0.01 | −1.75 | +0.02 | +0.00 |
| Water Dimer (S22-2) | −3.70 | +0.02 | +0.00 | −3.70 | +0.04 | +0.00 |
| Formic Acid Dimer (S22-3) | −13.89 | +0.10 | +0.07 | −13.83 | +0.03 | +0.04 |
| Formamide Dimer (S22-4) | −11.64 | +0.07 | +0.04 | −11.61 | +0.07 | +0.04 |
| Uracil Dimer HB (S22-5) | −16.11 | +0.17 | +0.11 | −16.10 | +0.22 | +0.21 |
| 2-Pyridone-2-Aminopyridine (S22-6) | −13.13 | +0.15 | +0.09 | −13.08 | +0.16 | +0.10 |
| Adenine-Thymine HB (S22-7) | −12.54 | +0.18 | +0.13 | −12.48 | +0.21 | +0.12 |
| Methane Dimer (S22-8) | −0.08 | +0.00 | +0.00 | −0.08 | +0.00 | +0.00 |
| Ethene Dimer (S22-9) | −0.26 | +0.02 | +0.01 | −0.25 | +0.02 | +0.01 |
| Benzene-Methane (S22-10) | −0.28 | +0.03 | +0.01 | −0.28 | +0.03 | +0.02 |
| Benzene Dimer PD (S22-11) | +0.16 | +0.12 | +0.05 | +0.17 | +0.23 | +0.18 |
| Pyrazine Dimer (S22-12) | −1.03 | +0.13 | +0.07 | −1.02 | +0.26 | +0.20 |
| Uracil Stack (S22-13) | −4.83 | +0.17 | +0.09 | −4.76 | +0.21 | +0.07 |
| Indole-Benzene Stack (S22-14) | −0.57 | +0.19 | +0.09 | −0.56 | +0.36 | +0.28 |
| Adenine-Thymine Stack (S22-15) | −5.46 | +0.25 | +0.13 | −5.39 | +0.28 | +0.14 |
| Ethene-Ethyne (S22-16) | −0.88 | +0.01 | +0.00 | −0.87 | +0.03 | +0.01 |
| Benzene-Water (S22-17) | −1.70 | +0.03 | +0.02 | −1.69 | +0.04 | +0.01 |
| Benzene-Ammonia (S22-18) | −0.94 | +0.03 | +0.02 | −0.93 | +0.02 | +0.01 |
| Benzene-HCN (S22-19) | −2.67 | +0.06 | +0.02 | −2.66 | +0.09 | +0.08 |
| Benzene Dimer T-Shape (S22-20) | −1.04 | +0.05 | +0.02 | −1.04 | +0.11 | +0.09 |
| Indole-Benzene T-Shape (S22-21) | −3.23 | +0.10 | +0.06 | −3.21 | +0.16 | +0.10 |
| Phenol Dimer (S22-22) | −4.66 | +0.09 | +0.05 | −4.64 | +0.10 | +0.05 |
Table II. Evaluating DLPNO-CCSD(T)-CCSD(T) and DLPNO-BCCD(T)-BCCD(T) dissociation energy errors on the complete set of S22 dimers, at the NormalPNO and TightPNO levels of convergence. Computations performed using the cc-pVDZ basis. All reported energetics are in kcal . and represent the DLPNO errors at normal and tight cutoffs with BCCD(T), and and represent DLPNO errors at normal and tight cutoffs with CCSD(T).
Figure 8. The RMSE, as well as maximum absolute errors of S22 dimers grouped by each interaction type (kcal mol), from the data in Table II for NormalPNO and TightPNO. (GPT-6 Astra assisted with making the plotting script, which is provided in the Supporting Information)
These results show that using Brueckner orbitals significantly reduces local correlation errors in both cases. At the tight cutoffs, local correlation error is reduced from an RMSE of to , with the most dramatic improvements in accuracy observed for the dispersion dominated (DD) complexes. The largest error of tight parameters is reduced from to with Brueckner orbitals. Similar trends are observed at normal cutoffs, with the RMSE reduced from to , and the max error reduced from to . One interesting observation is that the local correlation error associated with DLPNO-BCCD(T)/NormalPNO is similar to that at DLPNO-CCSD(T)/TightPNO, having similar values for MAE, RMSE, and max error. The virtual space also does not grow significantly between non-Brueckner and Brueckner computations, with the average PNO rank only increasing by 1-2 NOs. Thus, we believe that DLPNO-BCCD(T)/NormalPNO can represent a more economical alternative to DLPNO-CCSD(T)/TightPNO for larger systems.
Larger Systems
Having established the accuracy of DLPNO-BCCD(T)/NormalPNO in the previous section on non-covalent interactions, and the overall best reliability of DLPNO-BCCD(T) relative to CCSDT, we evaluated our DLPNO-BCCD(T) algorithm at normal cutoffs for two representative large dimers, coronene dimer (CCPD) and stacked octadecane dimer (CBH), in the L7 test set of Sedlak, Pulay, Hobza, and coworkers. [85] These systems are notable for being central to the ongoing controversy between CCSD(T) and FN-DMC. [104, 105, 106, 107] Results are presented in Table III. The differences between DLPNO-CCSD(T) and DLPNO-CCSD(T), as well as DLPNO-BCCD(T) and DLPNO-BCCD(T), are all within 0.1 kcal mol. However, the differences between DLPNO-BCCD(T) and DLPNO-CCSD(T), as well as their variants, are more significant, being around 0.2 kcal mol for CBH, and 0.6 kcal mol for CCPD. Based on our studies of the smaller S22 dimer interactions, we believe that this difference is attributed to the reduction of rank error associated with the local Brueckner ansatz, as systems with delocalized electrons, seen in systems like coronene dimer, are known to have higher errors, as seen in our S22 tests, compared to a dimer of an extended alkane like octadecane. We note that since canonical CCSD(T) and BCCD(T) are computationally intractable for these L7 systems, this is not rigorously established based on our data, and a more comprehensive benchmarking, though outside the scope of this work, is requisite before any firm conclusions can be drawn.
| Interaction | (nbf) | ||||
| Coronene Dimer PD () | 72 (792) | ||||
| Stacked Octadecane Dimer (CBH) | 112 (884) |
Table III. Counterpoise-corrected interaction energies of coronene dimer () and octadecane dimer (CBH) from the L7 [85] set, for DLPNO-CCSD(T), DLPNO-CCSD(T), DLPNO-BCCD(T), and DLPNO-BCCD(T). The number of atoms () and basis functions (nbf) are also included. All energetics are reported in kcal mol.
Conclusion
In this work, we have provided a scheme to compute Brueckner orbitals in the framework of domain-based local pair natural orbital (DLPNO) computations, resulting in the introduction of the DLPNO-BCCD(T) and DLPNO-BCCD(T) methods. These methods represent robust and reliable means to applying gold-standard quantum chemistry to larger systems, with DLPNO-BCCD(T) providing relative energies closest to CCSDT compared to the other non-iterative triples methods evaluated, on the order of or better, as shown in our tests on semibullvalene and bullvalene. A bonus benefit effect of using Brueckner orbitals is that it also reduces PNO errors relative to the canonical method, which we hypothesize arises from maximum overlap condition, allowing pair natural orbitals (PNOs) to be a more “exact” representation of the true pair electron density. With our tests on the S22 set of non-covalent dimers, we show that DLPNO-BCCD(T) produces similar errors with regards to dimer interaction energies at default cutoffs compared to DLPNO-CCSD(T) at tight cutoffs. This method can easily be applied to larger molecules, such as L7 non-covalent dimers. In the future, we plan to extend this algorithm for open-shell systems, as well as extend the Brueckner capabilities in our closed-shell code to higher orders of coupled-cluster theory (e.g., BCCDT for a -relaxed Brueckner reference). Thus, there is a strong case to be made for DLPNO-BCCD(T) and DLPNO-BCCD(T) to be utilized for routine quantum chemistry computations involving larger molecules.
Supporting Information
Additional figures (PDF). Geometries for systems studied, as well as full Psi4 and Molpro outputs (ZIP). Absolute energies (CSV). Relevant AI prompts (TXT).
Conflict of Interests
The authors have no conflicts of interest to declare.
Data Availability
The data that support the findings of this study are available with the article and its supporting information. The Psi4 source code for the implementations of DLPNO-BCCD(T)/(T) used in this work can be found at https://github.com/psi4/psi4/commit/46822f951c9bb6bf625537ac3456bacebd4ac9ec and https://github.com/andyj10224/psi4/tree/brueckner_lambda
Acknowledgements
A.J. acknowledges the US National Science Foundation (NSF) for a Graduate Research Fellowship (NSF Award 2236869). E.C.T. and H.F.S. acknowledge support from the US Department of Energy (DOE), Office of Science, Office of Basic Energy Sciences (BES) under Contract No. DE-SC0018412 as well as the US National Science Foundation under Grant No. 2410879.
T.Gy. and G.C. thank the National Research, Development and Innovation Office-NKFIH, K-146759; and the Momentum (Lendület) Program of the Hungarian Academy of Sciences for financial support, and acknowledge KIFÜ (Governmental Agency for IT Development, Hungary, https://ror.org/01s0v4q65) for awarding them access to the Komondor HPC facility based in Hungary.
This research was supported in part by resources and technical expertise from the Georgia Advanced Computing Resource Center, a partnership between the University of Georgia’s Office of the Vice President for Research and Office of the Vice President for Information Technology.
OpenAI ChatGPT (5.6 Sol and 6 Astra) and Anthropic Claude (5 Opus) were used to assist with parts of the software-development, literature search for the bibliography, the preparation of Figure 1 and Figure 8 in this manuscript, as well as for providing feedback during the drafting of this manuscript. All AI-generated content has been rigorously reviewed and verified by the human authors.
References
- [1]Crawford, T. D.; Schaefer III, H. F. In Reviews in Computational Chemistry; Lipkowitz, K., Boyd, D., Eds.; John Wiley & Sons, Ltd, 2000; Vol. 14; pp 33–136.
- [2]Bartlett, R. J.; Musial, M. Coupled-cluster theory in quantum chemistry. Rev. Mod. Phys. 2007, 79, 291–352.DOI
- [3]Knowles, P. J.; Handy, N. C. A new determinant-based full configuration interaction method. Chem. Phys. Lett. 1984, 111, 315–321.DOI
- [4]Sherrill, C. D.; Schaefer, H. F. In Advances in Quantum Chemistry; Löwdin, P.-O., Sabin, J. R., Zerner, M. C., Brändas, E., Eds.; Academic Press, 1999; Vol. 34; pp 143–269.
- [5]Helgaker, T.; Jørgensen, P.; Olsen, J. Molecular Electronic-Structure Theory; John Wiley & Sons: Chichester, 2000; Chapter 11, pp 523–597.DOI
- [6]Harvey, J. N.; Aschi, M. Modelling spin-forbidden reactions: recombination of carbon monoxide with iron tetracarbonyl. Faraday Discuss. 2003, 124, 129–143.
- [7]Benedek, Z.; Timár, P.; Szilvási, T.; Barcza, G. Sensitivity of coupled cluster electronic properties on the reference determinant: Can Kohn–Sham orbitals be more beneficial than Hartree–Fock orbitals? Journal of Computational Chemistry 2022, 43, 2103–2120.
- [8]Zamani, A. Y.; Zulueta, B.; Ricciuti, A. M.; Keith, J. A.; Carter-Fenk, K. Improving coupled cluster theory for strongly correlated molecules with Kohn–Sham density encoding. Nat. Commun. 2026, 17, 10791.
- [9]Purvis, G. D., III; Bartlett, R. J. A full coupled-cluster singles and doubles model: The inclusion of disconnected triples. J. Chem. Phys. 1982, 76, 1910–1918.
- [10]Hoffmann, M. R.; Schaefer, H. F. In Advances in Quantum Chemistry; Löwdin, P.-O., Ed.; Academic Press, 1986; Vol. 18; pp 207–279.
- [11]Noga, J.; Bartlett, R. J. The full CCSDT model for molecular electronic structure. J. Chem. Phys. 1987, 86, 7041–7050.DOI
- [12]Noga, J.; Bartlett, R. J. Erratum: The full CCSDT model for molecular electronic structure [J. Chem. Phys. 86, 7041 (1987)]. J. Chem. Phys. 1988, 89, 3401.
- [13]Scuseria, G. E.; Schaefer, H. F. A new implementation of the full CCSDT model for molecular electronic structure. Chem. Phys. Lett. 1988, 152, 382–386.DOI
- [14]Olsen, J.; Jørgensen, P.; Koch, H.; Balkova, A.; Bartlett, R. J. Full configuration–interaction and state of the art correlation calculations on water in a valence double-zeta basis with polarization functions. J. Chem. Phys. 1996, 104, 8007–8015.DOI
- [15]Larsen, H.; Olsen, J.; Jørgensen, P.; Christiansen, O. Full configuration interaction benchmarking of coupled-cluster models for the lowest singlet energy surfaces of N₂. J. Chem. Phys. 2000, 113, 6677–6686.DOI
- [16]Ruden, T. A.; Helgaker, T.; Jørgensen, P.; Olsen, J. Coupled-Cluster Connected-Quadruples Corrections to Atomization Energies. Chem. Phys. Lett. 2003, 371, 62–67.DOI
- [17]Tajti, A.; Szalay, P. G.; Császár, A. G.; Kállay, M.; Gauss, J.; Valeev, E. F.; Flowers, B. A.; Vázquez, J.; Stanton, J. F. HEAT: High accuracy extrapolated ab initio thermochemistry. J. Chem. Phys. 2004, 121, 11599–11613.
- [18]Karton, A.; Sylvetsky, N.; Martin, J. M. L. W4-17: A diverse and high-confidence dataset of atomization energies for benchmarking high-level electronic structure methods. J. Comput. Chem. 2017, 38, 2063–2075.DOI
- [19]Fishman, V.; Lőrincz, B. D.; Semidalas, E.; Barman, A.; Martin, J. M. L.; Nagy, P. R.; Kállay, M. Development of Local Natural Orbital Arbitrary Order Coupled Cluster Methods and Assessment through Connected Quadruples. J. Phys. Chem. A 2026, 130, 3233–3248.DOI
- [20]Karton, A. Highly Accurate CCSDT(Q)/CBS Reaction Barrier Heights for a Diverse Set of Transition Structures: Basis Set Convergence and Cost-Effective Approaches for Estimating Post-CCSD(T) Contributions. J. Phys. Chem. A 2019, 123, 6720–6732.DOI
- [21]Šimová, L.; Řezáč, J.; Hobza, P. Convergence of the Interaction Energies in Noncovalent Complexes in the Coupled-Cluster Methods Up to Full Configuration Interaction. J. Chem. Theory Comput. 2013, 9, 3420–3428.DOI
- [22]Řezáč, J.; Hobza, P. Describing Noncovalent Interactions beyond the Common Approxima- tions: How Accurate Is the “Gold Standard,” CCSD(T) at the Complete Basis Set Limit? J. Chem. Theory Comput. 2013, 9, 2151–2155.
- [23]Smith, D. G. A.; Jankowski, P.; Slawik, M.; Witek, H. A.; Patkowski, K. Basis Set Convergence of the Post-CCSD(T) Contribution to Noncovalent Interaction Energies. J. Chem. Theory Comput. 2014, 10, 3140–3150.DOI
- [24]Semidalas, E.; Boese, A. D.; Martin, J. M. L. Post-CCSD(T) corrections in the S66 noncovalent interactions benchmark. Chem. Phys. Lett. 2025, 863, 141874.DOI
- [25]Peterson, K. A.; Dunning, T. H. J. Intrinsic Errors in Several ab Initio Methods: The Dissociation Energy of N₂. J. Phys. Chem. 1995, 99, 3898–3901.DOI
- [26]Bak, K. L.; Jørgensen, P.; Olsen, J.; Helgaker, T.; Klopper, W. Accuracy of atomization energies and reaction enthalpies in standard and extrapolated electronic wave function/basis set calculations. J. Chem. Phys. 2000, 112, 9229–9242.DOI
- [27]Tasi, D. A.; Győri, T.; Czakó, G. On the development of a gold-standard potential energy surface for the OH⁻ + CH₃I reaction. Phys. Chem. Chem. Phys. 2020, 22, 3775–3778.DOI
- [28]Masios, N.; Irmler, A.; Schäfer, T.; Grüneis, A. Averting the Infrared Catastrophe in the Gold Standard of Quantum Chemistry. Phys. Rev. Lett. 2023, 131, 186401.DOI
- [29]Bak, K. L.; Jørgensen, P.; Olsen, J.; Helgaker, T.; Gauss, J. Coupled-cluster singles, doubles and triples (CCSDT) calculations of atomization energies. Chem. Phys. Lett. 2000, 317, 116–122.DOI
- [30]Karton, A.; Rabinovich, E.; Martin, J. M. L.; Ruscic, B. W4 theory for computational thermochemistry: In pursuit of confident sub-kJ/mol predictions. J. Chem. Phys. 2006, 125, 144108.pubmed.ncbi.nlm.nih.gov/17042580
- [31]Řezáč, J.; Šimová, L.; Hobza, P. CCSD[T] Describes Noncovalent Interactions Better than the CCSD(T), CCSD(TQ), and CCSDT Methods. J. Chem. Theory Comput. 2013, 9, 364–369.DOI
- [32]Spiegel, M.; Semidalas, E.; Martin, J. M. L.; Bentley, M. R.; Stanton, J. F. Post-CCSD(T) corrections to bond distances and vibrational frequencies: the power of Λ. Mol. Phys. 2024, 122, e2252114.DOI
- [33]Jones, G. H.; Semidalas, E.; Martin, J. M. L. How ‘de facto variational’ are fully iterative, approximate iterative, and quasiperturbative coupled cluster methods near equilibrium geometries? Pure Appl. Chem. 2025, 97, 1383–1392.DOI
- [34]Bomble, Y. J.; Vázquez, J.; Kállay, M.; Michauk, C.; Szalay, P. G.; Császár, A. G.; Gauss, J; Stanton, J. F. High-accuracy extrapolated ab initio thermochemistry. II. Minor improvements to the protocol and a vital simplification. J. Chem. Phys. 2006, 125, 064108.pubmed.ncbi.nlm.nih.gov/16942274
- [35]Harding, M. E.; Vázquez, J.; Ruscic, B.; Wilson, A. K.; Gauss, J.; Stanton, J. F. High-accuracy extrapolated ab initio thermochemistry. III. Additional improvements and overview. J. Chem. Phys. 2008, 128, 114111.DOI
- [36]Karton, A.; Taylor, P. R.; Martin, J. M. L. Basis set convergence of post-CCSD contributions to molecular atomization energies. J. Chem. Phys. 2007, 127, 064104.DOI
- [37]Klippenstein, S. J.; Harding, L. B.; Ruscic, B. Ab Initio Computations and Active Thermochemical Tables Hand in Hand: Heats of Formation of Core Combustion Species. J. Phys. Chem. A 2017, 121, 6580–6602.DOI
- [38]Elliott, S. N.; Keçeli, M.; Ghosh, M. K.; Somers, K. P.; Curran, H. J.; Klippenstein, S. J. High-Accuracy Heats of Formation for Alkane Oxidation: From Small to Large via the Automated CBH-ANL Method. J. Phys. Chem. A 2023, 127, 1512–1531.DOI
- [39]Karton, A.; Martin, J. M. L. Performance of W4 theory for spectroscopic constants and electrical properties of small molecules. J. Chem. Phys. 2010, 133, 144102.DOI
- [40]Puzzarini, C. Accurate molecular structures of small- and medium-sized molecules. Int. J. Quantum Chem. 2016, 116, 1513–1519.DOI
- [41]Puzzarini, C. Astronomical complex organic molecules: Quantum chemistry meets rotational spectroscopy. Int. J. Quantum Chem. 2017, 117, 129–138.DOI
- [42]Puzzarini, C.; Barone, V. Diving for Accurate Structures in the Ocean of Molecular Systems with the Help of Spectroscopy and Quantum Chemistry. Acc. Chem. Res. 2018, 51, 548–556.DOI
- [43]Puzzarini, C.; Stanton, J. F. Connections between the accuracy of rotational constants and equilibrium molecular structures. Phys. Chem. Chem. Phys. 2023, 25, 1421–1429.DOI
- [44]Puzzarini, C.; Ye, H.; Alessandrini, S. Isomerism of CH₂SO: Accurate structural, energetic, and spectroscopic characterization. J. Comput. Chem. 2024, 45, 777–786.DOI
- [45]Morgan, W. J.; Matthews, D. A.; Ringholm, M.; Agarwal, J.; Gong, J. Z.; Ruud, K.; Allen, W. D.; Stanton, J. F.; Schaefer, H. F. I. Geometric Energy Derivatives at the Complete Basis Set Limit: Application to the Equilibrium Structure and Molecular Force Field of Formaldehyde. J. Chem. Theory Comput. 2018, 14, 1333–1350.DOI
- [46]Allen, A. M.; Sargent, A. C.; Schaefer, H. F. Methylene: a turning point in the history of quantum chemistry and an enduring paradigm. Pure Appl. Chem. 2025, 97, 1621–1634.DOI
- [47]Gauss, J.; Stanton, J. F. Perturbative treatment of triple excitations in coupled-cluster calculations of nuclear magnetic shielding constants. J. Chem. Phys. 1996, 104, 2574–2583.DOI
- [48]Rizzo, A.; Gauss, J. Shielding polarizabilities calculated at the coupled-cluster singles and doubles level augmented by a perturbative treatment of triple excitations. J. Chem. Phys. 2002, 116, 869–877.DOI
- [49]Gauss, J. Analytic second derivatives for the full coupled-cluster singles, doubles, and triples model: Nuclear magnetic shielding constants for BH, HF, CO, N₂, N₂O, and O₃. J. Chem. Phys. 2002, 116, 4773–4776.DOI
- [50]Dubecký, M.; Jurečka, P.; Derian, R.; Hobza, P.; Otyepka, M.; Mitas, L. Quantum Monte Carlo Methods Describe Noncovalent Interactions with Subchemical Accuracy. J. Chem. Theory Comput. 2013, 9, 4287–4292.DOI
- [51]Dubecký, M.; Derian, R.; Jurečka, P.; Mitas, L.; Hobza, P.; Otyepka, M. Quantum Monte Carlo for noncovalent interactions: an efficient protocol attaining benchmark accuracy. Phys. Chem. Chem. Phys. 2014, 16, 20915–20923.DOI
- [52]Řezáč, J.; Dubecký, M.; Jurečka, P.; Hobza, P. Extensions and applications of the A24 data set of accurate interaction energies. Phys. Chem. Chem. Phys. 2015, 17, 19268–19277.DOI
- [53]Kowalski, K.; Piecuch, P. The method of moments of coupled-cluster equations and the renormalized CCSD[T], CCSD(T), CCSD(TQ), and CCSDT(Q) approaches. J. Chem. Phys. 2000, 113, 18–35.DOI
- [54]Shirley, W. A.; Petersson, G. The convergence of coupled-cluster methods for Be₂. Chem. Phys. Lett. 1991, 181, 588–590.DOI
- [55]Kucharski, S. A.; Bartlett, R. J. Connected quadruples for the frequencies of O₃. J. Chem. Phys. 1999, 110, 8233–8235.DOI
- [56]Feller, D.; Peterson, K. A.; Dixon, D. A. Further benchmarks of a composite, convergent, statistically calibrated coupled-cluster-based approach for thermochemical and spectroscopic studies. Mol. Phys. 2012, 110, 2381–2399.DOI
- [57]Martin, J. M. The eight-valence-electron systems re-examined: convergence of the coupled-cluster series and performance of quasiperturbative methods for quadruple excitations. Mol. Phys. 2014, 112, 785–793.DOI
- [58]Karton, A. Post-CCSD(T) contributions to total atomization energies in multireference systems. J. Chem. Phys. 2018, 149, 034102.DOI
- [59]Karton, A. Basis set convergence of high-order coupled cluster methods up to CCSDTQ567 for a highly multireference molecule. Chem. Phys. Lett. 2019, 737, 136810.DOI
- [60]Kucharski, S. A.; Bartlett, R. J. Noniterative energy corrections through fifth-order to the coupled cluster singles and doubles method. J. Chem. Phys. 1998, 108, 5243–5254.DOI
- [61]Crawford, T. D.; Stanton, J. F. Investigation of an asymmetric triple-excitation correction for coupled-cluster energies. Int. J. Quantum Chem. 1998, 70, 601–611.doi.org/10.1002/(sici)1097-461x(1998)70:4/5<601::aid-qua6>3.0.co;2-z
- [62]Taube, A. G.; Bartlett, R. J. Improving upon CCSD(T): ΛCCSD(T). I. Potential energy surfaces. J. Chem. Phys. 2008, 128, 044110.DOI
- [63]Taube, A. G.; Bartlett, R. J. Improving upon CCSD (T): ΛCCSD (T). II. Stationary formulation and derivatives. J. Chem. Phys. 2008, 128.DOI
- [64]Bozkaya, U. A noniterative asymmetric triple excitation correction for the density-fitted coupled-cluster singles and doubles method: Preliminary applications. J. Chem. Phys. 2016, 144.DOI
- [65]Bozkaya, U.; Schaefer, H. F., III Symmetric and asymmetric triple excitation corrections for the orbital-optimized coupled-cluster doubles method: Improving upon CCSD(T) and CCSD(T)Λ: Preliminary application. J. Chem. Phys. 2012, 136, 204114.DOI
- [66]Alagöz, Y.; Ünal, A.; Bozkaya, U. Efficient implementations of the symmetric and asymmetric triple excitation corrections for the orbital-optimized coupled-cluster doubles method with the density-fitting approximation. J. Chem. Phys. 2021, 155.DOI
- [67]Stanton, J. F. Why CCSD(T) works: a different perspective. Chem. Phys. Lett. 1997, 281, 130–134.DOI
- [68]Weflen, K. E.; Bentley, M. R.; Thorpe, J. H.; Franke, P. R.; Martin, J. M. L.; Matthews, D. A.; Stanton, J. F. Exploiting a Shortcoming of Coupled-Cluster Theory: The Extent of Non-Hermiticity as a Diagnostic Indicator of Computational Accuracy. The Journal of Physical Chemistry Letters 2025, 16, 5121–5127.DOI
- [69]Scuseria, G. E.; Schaefer, H. F. The optimization of molecular orbitals for coupled cluster wavefunctions. Chem. Phys. Lett. 1987, 142, 354–358.DOI
- [70]Sherrill, C. D.; Krylov, A. I.; Byrd, E. F. C.; Head-Gordon, M. Energies and analytic gradients for a coupled-cluster doubles model using variational Brueckner orbitals: Application to symmetry breaking in O₄⁺. J. Chem. Phys. 1998, 109, 4171–4181.DOI
- [71]Bozkaya, U.; Turney, J. M.; Yamaguchi, Y.; Schaefer, H. F.; Sherrill, C. D. Quadratically convergent algorithm for orbital optimization in the orbital-optimized coupled-cluster doubles method and in orbital-optimized second-order Møller–Plesset perturbation theory. J. Chem. Phys. 2011, 135, 104103.
- [72]Robinson, J. B.; Knowles, P. J. Quasi-variational coupled cluster theory. J. Chem. Phys. 2012, 136, 054114.DOI
- [73]Thouless, D. J. Stability conditions and nuclear rotations in the Hartree-Fock theory. Nucl. Phys. 1960, 21, 225–232.DOI
- [74]Nesbet, R. K. Brueckner’s Theory and the Method of Superposition of Configurations. Phys. Rev. 1958, 109, 1632–1638.DOI
- [75]Paldus, J.; Čížek, J.; Keating, B. A. Stability Conditions for Maximum-Overlap (Brueckner) Independent-Particle Wave Functions. Phys. Rev. A 1973, 8, 640–649.DOI
- [76]Chiles, R. A.; Dykstra, C. E. An electron pair operator approach to coupled cluster wave functions. Application to He₂, Be₂, and Mg₂ and comparison with CEPA methods. J. Chem. Phys. 1981, 74, 4544–4556.DOI
- [77]Handy, N. C.; Pople, J. A.; Head-Gordon, M.; Raghavachari, K.; Trucks, G. W. Size-consistent Brueckner theory limited to double substitutions. Chem. Phys. Lett. 1989, 164, 185–192.DOI
- [78]Raghavachari, K.; Pople, J. A.; Replogle, E. S.; Head-Gordon, M.; Handy, N. C. Size-consistent Brueckner theory limited to double and triple substitutions. Chem. Phys. Lett. 1990, 167, 115–121.DOI
- [79]Watts, J. D.; Gauss, J.; Bartlett, R. J. Coupled-cluster methods with noniterative triple excitations for restricted open-shell Hartree–Fock and other general single determinant reference functions. Energies and analytical gradients. J. Chem. Phys. 1993, 98, 8718–8733.DOI
- [80]Tew, D. P. Explicitly correlated coupled-cluster theory with Brueckner orbitals. J. Chem. Phys. 2016, 145.DOI
- [81]Robinson, J. B.; Knowles, P. J. Benchmark Quasi-Variational Coupled Cluster Calculations of Multiple Bond Breaking. J. Chem. Theory Comput. 2012, 8, 2653–2660.DOI
- [82]Taube, A. G. Developments in Coupled-Cluster Theory Gradients and Potential Energy Surfaces. Ph.D. dissertation, University of Florida, 2008.
- [83]Toth, E. C.; Jiang, A.; Schaefer, H. F., III Linear-Scaling Asymmetric Triples Correction through the Solution of the DLPNO–CCSD Lambda Equations: DLPNO–CCSD(T)Λ. J. Chem. Theory Comput. 2026, 22, 7667–7681.
- [84]Jurečka, P.; Šponer, J.; Černý, J.; Hobza, P. Benchmark database of accurate (MP2 and CCSD(T) complete basis set limit) interaction energies of small model complexes, DNA base pairs, and amino acid pairs. Phys. Chem. Chem. Phys. 2006, 8, 1985–1993.pubmed.ncbi.nlm.nih.gov/16633685
- [85]Sedlak, R.; Janowski, T.; Pitoňák, M.; Řezáč, J.; Pulay, P.; Hobza, P. Accuracy of Quantum Chemical Methods for Large Noncovalent Complexes. J. Chem. Theory Comput. 2013, 9, 3364–3374.pubmed.ncbi.nlm.nih.gov/24098094
- [86]Jiang, A.; Glick, Z. L.; Poole, D.; Turney, J. M.; Sherrill, C. D.; Schaefer, H. F. Accurate and efficient open-source implementation of domain-based local pair natural orbital (DLPNO) coupled-cluster theory using a t₁-transformed Hamiltonian. J. Chem. Phys. 2024, 161, 082502.
- [87]Jiang, A.; Schaefer, H. F.; Turney, J. M. Linear-Scaling Local Natural Orbital-Based Full Triples Treatment in Coupled-Cluster Theory. J. Chem. Theory Comput. 2025, 21, 2386–2401.DOI
- [88]Pulay, P. Convergence Acceleration of Iterative Sequences. The Case of SCF Iteration. Chem. Phys. Lett. 1980, 73, 393–398.DOI
- [89]Schönemann, P. H. A Generalized Solution of the Orthogonal Procrustes Problem. Psychometrika 1966, 31, 1–10.DOI
- [90]Kuhn, H. W. The Hungarian Method for the Assignment Problem. Naval Res. Logist. Q. 1955, 2, 83–97.
- [91]Sun, Q. Co-iterative Augmented Hessian Method for Orbital Optimization. 2016.arxiv.org/abs/1610.08423
- [92]Clement, M. C.; Wang, X.; Valeev, E. F. Robust Pipek–Mezey Orbital Localization in Periodic Solids. J. Chem. Theory Comput. 2021, 17, 7406–7415.DOI
- [93]Foster, J. M.; Boys, S. F. Canonical Configurational Interaction Procedure. Rev. Mod. Phys. 1960, 32, 300–302.
- [94]Pipek, J.; Mezey, P. G. A fast intrinsic localization procedure applicable for ab initio and semiempirical linear combination of atomic orbital wave functions. J. Chem. Phys. 1989, 90, 4916–4926.DOI
- [95]Lehtola, S.; Jónsson, H. Pipek–Mezey Orbital Localization Using Various Partial Charge Estimates. J. Chem. Theory Comput. 2014, 10, 642–649.DOI
- [96]Knizia, G. Intrinsic Atomic Orbitals: An Unbiased Bridge between Quantum Theory and Chemical Concepts. J. Chem. Theory Comput. 2013, 9, 4834–4843.
- [97]Werner, H.-J. et al. The Molpro quantum chemistry package. J. Chem. Phys. 2020, 152, 144107.DOI
- [98]Robinson, J. B.; Knowles, P. J. Breaking multiple covalent bonds with Hartree–Fock-based quantum chemistry: Quasi-Variational Coupled Cluster theory with perturbative treatment of triple excitations. Phys. Chem. Chem. Phys. 2012, 14, 6729–6732.DOI
- [99]Robinson, J. B.; Knowles, P. J. Rigorously extensive orbital-invariant renormalized perturbative triples corrections from quasi-variational coupled cluster theory. J. Chem. Phys. 2013, 138, 074104.DOI
- [100]Black, J. A.; Knowles, P. J. Quasi-variational coupled-cluster theory: Performance of perturbative treatments of connected triple excitations. J. Chem. Phys. 2018, 148, 194102.DOI
- [101]Black, J. A.; Knowles, P. J. Statistical analysis of activation and reaction energies with quasi-variational coupled-cluster theory. Mol. Phys. 2018, 116, 1421–1427.DOI
- [102]Győri, T.; Czakó, G. Automating the Development of High-Dimensional Reactive Potential Energy Surfaces with the robosurfer Program System. J. Chem. Theory Comput. 2020, 16, 51–66.DOI
- [103]Karton, A. Cope rearrangements in shapeshifting molecules re-examined by means of high-level CCSDT(Q) composite ab initio methods. Chem. Phys. Lett. 2020, 759, 138018.DOI
- [104]Al-Hamdani, Y. S.; Nagy, P. R.; Zen, A.; Barton, D.; Kállay, M.; Brandenburg, J. G.; Tkatchenko, A. Interactions between large molecules pose a puzzle for reference quantum mechanical methods. Nat. Commun. 2021, 12, 3927.
- [105]Villot, C.; Ballesteros, F.; Wang, D.; Lao, K. U. Coupled Cluster Benchmarking of Large Noncovalent Complexes in L7 and S12L as Well as the C60 Dimer, DNA–Ellipticine, and HIV–Indinavir. J. Phys. Chem. A 2022, 126, 4326–4341.DOI
- [106]Lambie, S.; Kats, D.; Usvyat, D.; Alavi, A. On the applicability of CCSD(T) for dispersion interactions in large conjugated systems. J. Chem. Phys. 2025, 162, 114112.
- [107]Schäfer, T.; Irmler, A.; Gallo, A.; Grüneis, A. Understanding discrepancies in noncovalent interaction energies from wavefunction theories for large molecules. Nat. Commun. 2025, 16, 9108.DOI






