WorksPhysics and Astronomy
ElemCo.jl: A Julia package for electron-correlation methods
Abstract
We present ElemCo.jl, an open-source Julia package for molecular electronic structure and properties calculations with a particular emphasis on Coupled Cluster and Distinguishable Cluster methods. The package provides a high-level, macro-based user interface which makes routine calculations accessible to users with no prior Julia experience. A newly developed visualizer, JLmol, assists in the graphical preparation of ElemCo.jl input files and displays results such as molecular orbitals. ElemCo.jl offers restricted closed-shell and unrestricted variants of state-of-the-art electron-correlation methods such as FCI, MP2, CCSD(T) as well as EOM-CCSD and the DC methods DCSD and DC-CCSDT. In addition to traditional approaches, ElemCo.jl provides recently developed methods that are currently unique to the package. These include tensor-decomposed implementations of DCSD and DC-CCSDT (SVD-DCSD and SVD-DC-CCSDT) as well as two-determinant and fixed-reference CC and DC methods. For excited-state calculations, ElemCo.jl also offers EOM-DCSD, which is benchmarked in this work against CC3 on the QUEST3 benchmark set, alongside EOM-CCSD. Furthermore, both ground and excited states, including those of multireference character, can be treated using CIPHI - an efficient selected Configuration Interaction (CI) approach employing a CIPSI/Heat-Bath-CI-based algorithm. Users can directly invoke internal functions from the input file, enabling them to test and compose new methods without the need to modify ElemCo.jl's source code. Also, ElemCo.jl can be readily interfaced with external quantum chemistry codes through the Fcidump format, for example as the high-level solver for transcorrelated Hamiltonians, periodic embedded fragments, etc. We provide a detailed overview of ElemCo.jl's methodological repertoire and discuss its technical details and implementation, performance, interfaces, and usage.
Introduction
Quantum chemistry program packages are nowadays widely used by both theoreticians and experimentalists. Among wavefunction-based electron-correlation methods, Coupled Cluster (CC) theory is widely regarded as the gold standard of molecular electronic-structure theory because it provides systematically improvable and size-extensive descriptions of dynamic electron correlation. [1, 2, 3] Thus, a large number of quantum chemistry program packages provide CC implementations, including but not limited to Molpro, CFOUR, Orca, Turbomole, MRCC, NWChem, and Gamess. [4, 5] [6] [7] [8] [9] [10] [11] These packages also offer a wide range of electron-correlation methods beyond CC, including approaches designed to describe systems with significant static electron correlation. These codes, typically written in Fortran or C/C++, represent decades of development and are often highly efficient. However, understanding their source code, adapting their functionality to specific needs, and extending them with new methods can be challenging, particularly for early-career scientists and experimental chemists. Moreover, when exploring new ideas for which only a proof of concept is initially required, the implementation process can be unnecessarily lengthy, even though high performance is not yet essential.
More recently, Python-based packages such as PySCF [12, 13] or PyBEST [14] have been developed with an emphasis on code readability and ease of extension. In these packages, quantum-chemical building blocks exposed as scriptable objects have considerably lowered the barrier to method development. Despite its productivity benefits, Python suffers from the “two-language problem”: performance-critical kernels are typically implemented in a compiled language, which creates a split between a user-friendly scripting layer and a less transparent but highly efficient core.
The Julia programming language[15] was designed to overcome this limitation: it combines the interactivity and expressiveness of a dynamic language with performance approaching that of C or Fortran, achieved through just-in-time compilation and an aggressive type-specializing compiler. Its multiple-dispatch paradigm makes it natural to write generic algorithms that operate transparently. These features make Julia an attractive choice for quantum chemistry applications. For example, the DFTK[16] and Fermi.jl[17] packages leverage the advantages offered by Julia.
ElemCo.jl[18] (pronounced “elemcoil”, logo in Figure 1) was originally developed to provide a coupled-cluster (CC) solver for non-Hermitian Hamiltonians arising in transcorrelation[19, 20, 21] and therefore not supported by most conventional CC codes. The initial implementation was limited to CCSD and Distinguishable Cluster Singles and Doubles (DCSD)[22, 23] for general Hamiltonians. Over time, however, ElemCo.jl evolved substantially owing to Julia’s combination of extensibility and high performance and has grown into a standalone quantum chemistry package. Consequently, a wide range of CC and DC methods for ground and excited states, together with mean-field, perturbative, and Configuration Interaction (CI) approaches, have been incorporated. Furthermore, numerical capabilities such as density fitting (DF)[24, 25, 26, 27, 28], Cholesky decomposition[29, 30, 31], Laplace transform[32, 33, 34], singular-value decomposition (SVD)[35], orbital localization[36, 37, 38, 39, 40], and the availability of Davidson[41] and direct inversion in the iterative subspace (DIIS) solvers[42, 43], among others, make ElemCo.jl a versatile platform for implementing electron-correlation methods with state-of-the-art performance.
Figure 1. The ElemCo.jl logo.
Overall, the result is a package that is not only easy to extend but also straightforward to use. Calculations are initiated through intuitive macros and can be embedded in ordinary Julia scripts, such that a complete input file typically consists of only a handful of lines and requires no knowledge of Julia beyond copying a template. At the same time, because every method is implemented in pure Julia, users can directly invoke internal functions from the input file, enabling them to test and compose new methods without the need to modify ElemCo.jl’s source code.
It should be emphasized that, beyond providing accessible and well-established methods, ElemCo.jl includes a variety of recently developed approaches that are currently unique to the package, including the tensor-decomposed SVD-DCSD and SVD-DCCSDT methods[44, 45], the two-determinant DCSD (2D-DCSD), fixed-reference DCSD (FR-DCSD) and DC-CCSDT (FR-DC-CCSDT) methods[46, 47, 48], and the CIPHI selected CI method. Equally distinctive is the systematic support for non-Hermitian Hamiltonians, including ACCSD(T), which can be employed, for example, in transcorrelation. Making both these newly developed methods and established approaches freely and openly available constitutes a central motivation of the project.
In the remainder of this paper, we describe the design and implementation of ElemCo.jl (§ 1.1), survey the available methods (Sec. III), the interfaces (Sec. IV), illustrate usage (§ 1.4 and VII), the performance (§ 1.5), and close with an outlook on ongoing developments (Sec. VIII).
Design and implementation details
Installation and user interface
ElemCo.jl is written entirely in Julia (version ) and can be installed with a single command in the Julia package manager, namely Pkg.add("ElemCo"). The only external numerical dependency is the LIBCINT library,[49] for integral evaluation, which is distributed as a precompiled binary and therefore requires no manual compilation by the user. To ensure that calculations can be performed without any expertise in Julia, the user interface is built around macros of the form @macroname. An input file is an ordinary Julia script that loads the package via using ElemCo and then executes a sequence of macro calls. For example, the following input fully specifies a Hartree–Fock (HF) calculation followed by a CCSD(T) calculation for a nitrogen molecule with the geometry in the XYZ-format, using a valence double-zeta basis set:
using ElemCo
geometry = "2
N -0.556 0.000 0.000
N 0.556 0.000 0.000"
basis = "vdz"
@hf
@cc ccsd(t)
The macros (@hf, @cc, etc.) expand into ordinary function calls at parse time, so the convenience layer carries no run-time overhead and remains fully transparent. Because the input is itself a Julia program, control flow, loops over geometries, parameter scans, and post-processing can be added with the full power of the language when desired. The ElemCo.jl-dedicated visualizer JLmol[50] – which is essentially a wrapper around the JSMOL library[51, 52] – enables the graphical generation of ElemCo.jl input files and provides visualization capabilities, including the display of molecular orbitals. Available both as a desktop application and as a browser-based application, it offers a convenient alternative to manual input preparation.
Beyond the macros, every routine in ElemCo.jl is an ordinary, documented Julia function. Users seeking greater flexibility can bypass the macro interface entirely to script novel workflows or implement new methods by directly invoking the underlying functions, without any loss of capability. This “two-tier” design – a declarative macro layer over a fully exposed functional core – enables ElemCo.jl to serve both non-expert users and method developers from a single code base.
Detailed information on all macros, functions, and available options can be found on the ElemCo.jl website at https://elem.co.il and in our GitHub repository at https://github.com/fkfest/ElemCo.jl.
Macros, options and workflow
Table I provides an overview of the macros available in ElemCo.jl for invoking quantum chemical methods.
The macros rely on various local variables to specify the calculation. The system information, i.e., the molecular geometry, the basis set, or the Hamiltonian (as an FCIDUMP file[53]), is specified through geometry, basis, or fcidump variables. This information, as well as various settings (including some technical settings such as the scratch directory), is stored in an ECInfo object under the variable EC (which also lives in the local scope), which is created automatically when the first macro is called. The ECInfo object is passed implicitly to all subsequent macros, so that the user does not need to manage it explicitly. It can be reset by calling the @ECinit macro, which creates a new ECInfo object and discards the previous one.
Also the settings of the respective methods, such as convergence thresholds or the maximum number of iterations, are controlled through macros. Options are specified using the @set macro, either globally or, preferably, locally within a single calculation block, ensuring that they are automatically restored afterwards and cannot inadvertently affect subsequent steps, as illustrated below.
@cc ccsd begin
#number of maximum iterations and convergence threshold
@set cc maxit=100 thr=1.e-10
end
The settings which are saved in the EC::ECInfo object are organized into thematic groups, among them wf for general wavefunction settings, scf for self-consistent field settings, cc for CC or DC settings, eom for EOM settings and so on.
Aside from full quantum chemical calculations and their associated settings, ElemCo.jl also provides dedicated macros for tasks such as evaluating integrals, defining dummy atoms, freezing orbitals, and performing orbital rotations. In addition, workflows can chain multiple macros together, with intermediate results such as orbitals, integrals, and amplitudes passed implicitly through a scratch directory and saved to or restored from wavefunction dumps using the @savewf, @loadwf, and @copywf macros.
Tensor engine and storage
The majority of electronic-structure methods rely on tensor contractions, and their computational efficiency and scalability are heavily determined by how efficiently these contractions can be performed. In ElemCo.jl, these contractions are expressed through a concise Einstein-summation macro, called @mtensor, layered on TensorOperations.jl[54], which dispatches binary contractions to optimized BLAS routines (e.g., via MKL). Array views and reusable memory buffers (Buffers.jl) reduce allocations in the hot loops. Large intermediates that do not fit in memory can be memory-mapped to disk transparently, so that the same code path serves both in-core and out-of-core regimes; amplitude and other tensors are written to and read from the scratch directory with save!/load!/mmap helpers. Cluster amplitudes follow a virtual-first index ordering—singles as T1[a,i] (T_i^a) and doubles as T2[a,b,i,j] (T_{ij}^{ab})—so that the rate-determining contractions map directly onto efficient BLAS calls.
A key consequence of Julia’s parametric type system is that the entire correlation machinery is written generically in the element type. The same source code therefore runs with double-precision real (Float64) or complex (ComplexF64) arithmetic: when a complex FCIDUMP is loaded, the element type propagates automatically through the solvers, tensor tools, and whole quantum chemical algorithms; and the just-in-time compilation and multiple dispatch ensure that the support for complex-valued Hamiltonians does not incur any overhead.
Integrals
The one- and two-electron integrals in ElemCo.jl are either computed through its interface to LIBCINT[48] or can be read from an FCIDUMP file, a wide spread format for many-body Hamiltonians.[53] The latter option allows easy interfacing with other quantum chemistry software, also for non-symmetric Hamiltonians such as transcorrelated ones. ElemCo.jl uses the density fitting (DF) approximation[24, 25, 26, 27, 28] for LIBCINT-generated integrals throughout, but the integrals are usually assembled. From the version 0.16 (to be released soon), also standard, non-density-fitted integrals are available. A library of standard orbital and auxiliary, density-fitting (JKFIT and MPFIT) basis sets is included in ElemCo.jl, obtained from the Basis Set Exchange library[55].
The lifecycle of the two-electron integrals follows their origin. FCIDUMP integrals are in the molecular-orbital (MO) basis and are treated as user-provided data: once read, they remain the active integral set for all subsequent correlated calculations until they are replaced, e.g. by pointing the fcidump variable to a different file. On the DF route the MO integrals are, by contrast, a transient object: the three-index intermediates are computed via LIBCINT and the required four-index MO blocks are assembled on demand inside the correlated calculation and deleted when it finishes. If several calculations are to share one DF-MO integral set, it can be generated explicitly with @dfints; an explicitly generated set persists like a Fcidump one. Finally, the exact (non-DF) AO integrals are generated automatically by the Hartree-Fock macros @hf/@uhf, or on explicit request with @ints (or by setting the int option df=false), and persist in the scratch directory for subsequent correlated calculations until the geometry or basis set is changed. Many correlated methods then run AO-direct: MP2, CCSD/DCSD and their variants, perturbative triples, equations and properties, EOM, and SVD-DC-CCSDT contract the AO integrals directly, and no full set of MO integrals is ever formed. A conventional MO-basis calculation on top of the exact AO integrals remains available through @moints (or by setting the int option ao_direct=false), which transforms the integrals to the MO basis and stores them persistently, in the same packed representation as the Fcidump integrals.
| Class | Methods |
| Mean field | @hf, @uhf, @dfhf, @dfuhf, @bohf, @bouhf, @dfmcscf |
| Møller-Plesset perturbation theory | @dfmp2, @cc mp2, @dfcc mp2 |
| Coupled cluster methods | @cc [methodname], @dfcc [methodname] |
| Distinguishable cluster methods | @cc [methodname], @dfcc [methodname] |
| Configuration interaction | @fci, @ciphi |
| Equation of motion | @cc eom-[methodname] |
Table 1. Overview of the macros available in ElemCo.jl to call quantum chemical methods.
The exact four-index AO integrals are stored as “± supermatrices”. Real AO integrals possess the full eightfold permutational symmetry, which would in principle permit an packing; the store is, however, deliberately built on the smaller symmetry group that survives for complex and (to some degree) for similarity-transformed (e.g. transcorrelated) integrals: the joint exchange symmetry together with hermiticity (for non-transcorrelated integrals). By splitting the integrals into bra-exchange symmetric and antisymmetric combinations,
the joint symmetry is transformed into two independent ones: and are symmetric (hermitian, for complex integrals) matrices in the packed pair space, and both index pairs and become triangular. Only the lower block triangle of each supermatrix is kept on disk (i.e., the total storage requirement is ), with the upper part reconstructed as the (conjugate-)hermitian mirror at read time. The Hartree-Fock and the AO-direct correlated methods stream this store panel-wise through a memory map; details of the layout and of the AO-driven contractions are given in the ElemCo.jl theory documentation.
The molecular-orbital integrals are stored and addressed using physicists’ notation, . Conventional four-index integrals are kept in a memory-efficient packed representation: exploiting their permutational symmetry, the last two orbital indices are combined into a single upper-triangular composite index , such that the integrals are stored as a three-index array, roughly halving the storage requirements, utilizing the only symmetry of two-electron integrals from non-Hermitian Hamiltonians. Whenever a particular block is requested, it is expanded back to its full four-index form on the fly. By contrast, density-fitted and Cholesky integrals are stored directly in their factorized three-index representation, , where labels the (Coulomb-metric-orthogonalized) auxiliary functions.
These sub-blocks of the integrals are retrieved through a compact, string-based interface. The functions ints1(EC,"ov") and ints2(EC,"oovv") return the one- and two-electron integrals over the requested orbital subspaces, where each character names a space: o/v for occupied/virtual () orbitals, O/V for their counterparts, : for the full molecular-orbital space, plus further labels for doubly/singly occupied and positron spaces. The order of the characters is significant ("ov" "vo"), and the upper/lower case of the labels selects the spin case—so that ints2(EC,"oovv") returns the closed-shell block (indexed [i,j,a,b]), whereas ints2(EC,"ooVV") returns the corresponding block. The EC::ECInfo object carries the orbital-space definitions.
Quantum chemical methods
ElemCo.jl provides a broad range of electron-correlation methods. Table II summarizes the currently implemented approaches, which are described in detail in the following subsections. The working equations of most methods can be found in the documentation of ElemCo.jl.[18]
| Class | Methods |
| Mean field | HF, UHF, DF-HF, DF-UHF, BO-HF, BO-UHF, DF-MCSCF |
| Perturbation | DF-MP2, MP2, UMP2, SOS-LT-DF-MP2 |
| Coupled cluster | CCSD, RCCSD, UCCSD, CCSD(T), RCCSD(T), UCCSD(T), CCSD(T), UCCSD(T), CCSDT, UCCSDT |
| Distinguishable cluster | DCSD, RDCSD, UDCSD, DC-CCSDT, UDC-CCSDT |
| Quasi-variational CC and DC | QV-CCD, OQV-CCD, QV-DCD, OQV-DCD |
| Tensor decomposed | SVD-DCSD, SVD-DC-CCSDT |
| Two-determinant | 2D-CCSD, 2D-DCSD, FR-CCSD, FR-DCSD, FR-CCSDT, FR-DC-CCSDT |
| Configuration interaction | FCI, CIPHI |
| Excited states (EOM) | EOM-CCSD, EOM-RCCSD, EOM-UCCSD, EOM-DCSD, EOM-RDCSD, EOM-UDCSD |
| Tensor networks | DMRG (via ITensors.jl) |
Table II. TABLE II. Overview of the quantum chemical methods available in ElemCo.jl. Closed-shell (spin-restricted), restricted-open-shell (prefix R), and spin-unrestricted (prefix U) variants are available for the HF, MP2, CC, DC, and excited-state methods as indicated. All abbreviations can be found in Section III.
Hartree-Fock methods and orbital localization
ElemCo.jl implements standard and density-fitted Hartree-Fock in both spin-restricted closed-shell (HF and DF-HF) and spin-unrestricted (UHF and DF-UHF) variants. In addition to these canonical approaches, ElemCo.jl provides biorthogonal Hartree-Fock (BO-HF/BO-UHF), in which distinct left and right orbital sets are determined, a feature that is crucial for the orbital optimization in non-Hermitian Hamiltonians. For multireference problems, a density-fitted multiconfigurational self-consistent-field (DF-MCSCF) method employing combined first- and second-order orbital optimization[56, 57, 58] is also available.
Once the orbitals have been obtained, localization of the occupied orbitals is supported through the intrinsic bond orbital (IBO),[36] Pipek-Mezey,[37] and Foster-Boys[38] schemes, while orthogonalized projected atomic orbitals[39, 40] can be employed for the virtual space.
Møller-Plesset perturbation theory
Second-order Møller-Plesset perturbation theory is available in closed-shell (MP2, DF-MP2), and unrestricted (UMP2) variants. A scaled-opposite-spin Laplace-transform density-fitted formulation (SOS-LT-DF-MP2) is also provided for reduced-scaling estimates.
Coupled-cluster methods
Coupled Cluster (CC) methods available in ElemCo.jl include CC with Singles and Doubles (CCSD),[1] its perturbative triples extension CCSD(T),[2] the corresponding CCSD(T) variant,[59] and Coupled Cluster with iterative triples (CCSDT). All aforementioned methods are implemented in closed-shell, restricted open-shell (R),[60, 61] and spin-unrestricted (U) variants. For the implementation of CCSDT and UCCSDT, the working equations were automatically generated using our second-quantization program Quantwo,[62] which derives the working equations symbolically, factorizes the resulting contractions, and emits the corresponding Julia tensor-contraction kernels.
The CCSD factorization is partially based on Ref. 63, i.e., the most expensive step of the CCSD iteration is performed in the full (MO or AO) basis. In the AO-direct implementation, the most expensive contraction is performed in the AO basis, directly contracting the combinations of AO integrals with the combinations of amplitudes, which reduces the cost of the contraction by a factor of 2. The remaining contractions are performed in the MO basis.
Distinguishable Cluster methods
A central feature of ElemCo.jl is the availability of Distinguishable Cluster (DC) approximations[22, 23] to all implemented CC methods and beyond. The DC approximation is obtained from CCSD by removing quadratic exchange terms from the doubles-amplitude equations and rescaling a few other terms. This modification substantially improves the description of closed- and open-shell reactions, atomization energies, ionization potentials, electron affinities, interaction energies, equilibrium geometries, spectroscopic constants, systems with multireference character, and bond-breaking processes, while simultaneously restoring exactness for two-electron systems and preserving size extensivity, orbital-rotation invariance, and particle-hole symmetry.[64, 65, 66, 67]
ElemCo.jl provides DCSD together with its restricted open-shell and unrestricted variants (RDCSD and UDCSD), as well as iterative triples extensions[48, 68, 69] in the form of DC-CCSDT and UDC-CCSDT. As for the CCSDT and UCCSDT implementations, the corresponding working equations for their DC counterparts were generated automatically using our second-quantization program Quantwo.[62]
Quasi-Variational Coupled Cluster and Distinguishable Cluster methods
Quasi-variational coupled-cluster theory [70, 71] is a Hermitian variant of CC theory that incorporates some degree of variationality into the otherwise non-variational CC ansatz. Both quasi-variational coupled-cluster doubles (QVCCD) [70] and quasi-variational distinguishable cluster doubles (QVDCD) [72] are implemented in ElemCo.jl in their closed-shell formulations, together with their orbital-optimized variants, OQVCCD and OQVDCD.
Tensor-decomposed methods
To extend high-accuracy methods to larger systems, ElemCo.jl implements tensor-decomposed DC methods based on a singular-value decomposition (SVD) of the cluster amplitudes. For SVD-DC-CCSDT, [44] the triples amplitudes are represented in a compact SVD basis,
where is a small, fully symmetric three-index core tensor and the compression coefficients connect the SVD indices , , and to the occupied () and virtual () orbital spaces. The triples amplitudes are optimized directly in this SVD-basis. Combined with DF, this reduces the nominal computational scaling of DC-CCSDT from to , as the number of SVD functions is expected to grow linearly with system size. The SVD basis is constructed from an auxiliary CC3-type triples density, using a threshold to retain only the most important singular vectors based on the magnitude of the corresponding singular values. An optional hybrid correction, , accelerates convergence with respect to the SVD threshold. An analogous SVD-DCSD method is also implemented, in which the doubles amplitudes are represented in an SVD basis. Currently, SVD-DCSD is available as a pilot implementation and is scheduled for inclusion in the next ElemCo.jl release.
Two-determinant and fixed-reference methods
For challenging spin states in which single-determinant wavefunction ansätze become inadequate, ElemCo.jl offers two-determinant variants of CCSD and DCSD (2D-CCSD and 2D-DCSD), [46, 47] together with fixed-reference methods (FR-CCSD and FR-DCSD) and their triples extensions FR-CCSDT and FR-DC-CCSDT. [47, 48] The defining feature of the fixed-reference approaches is that a selected reference doubles amplitude remains fixed during the iterative optimization. These methods can be applied to biradical systems to enforce the correct spin state.
Configuration interaction: FCI and CIPHI
In addition to an efficient full configuration interaction (FCI) implementation, ElemCo.jl also offers its own selected CI method, termed CIPHI (Configuration Interaction using a Perturbative/Heat-bath Iterative selection), which becomes an effective alternative to FCI for active spaces that are too large for the latter. As its name suggests, CIPHI incorporates ideas from both CIPSI [73] and heat-bath CI (HCI). [74, 75] Unlike the hybrid CIPSI/HCI approach of Ref. [76], we employ a determinantal basis rather than configuration state functions, which allows for the perturbative selection criterion to be evaluated with negligible computational cost. CIPHI is similar in spirit to ASCI from Ref. [77], and differs primarily in the selection and prescreening algorithm, which we summarize below.
CIPHI iteratively selects the most important determinants for the CI expansion. Each macro-iteration consists of four steps: (i) the pre-selection of new determinants based on the heat-bath criterion, (ii) the final perturbative selection of new determinants, (iii) construction (extension) of the CI Hamiltonian in the selected determinant space, and (iv) diagonalization of the Hamiltonian in the enlarged determinant space. The procedure is repeated until either a target expansion size is reached or the CI energy has converged. At the end, a perturbative correction is evaluated.
(i) Heat-bath preselection. For every selected determinant with CI coefficient , singly and doubly excited candidate determinants satisfying the heat-bath criterion
are generated, where is a screening threshold (by default , with the master selection threshold). To make this step efficient, the double-excitation matrix elements are precomputed once and stored, for each ordered orbital pair, in lists sorted by decreasing (a precomputed-heat-bath, PCHB, data structure following Holmes et al. [74]). The sorted lists allow the generation loop to stop as soon as falls below the adaptive threshold , so that the cost of proposing candidates scales with the number of important connections rather than with the full size of the excitation manifold. An additional instantaneous coefficient threshold can be used to discard candidates with negligible instantaneous weight,
on the fly (this is however optional). The denominator in Equation (4) is efficiently obtained using the relation
where is the difference between the diagonal Hamiltonian matrix elements of the selected and candidate determinants, which only depends on the excitation information (and the diagonal of the Fock matrix of ), and therefore can be precomputed and stored along with the PCHB data structure.
The candidate determinants that survive the heat-bath preselection are stored as a dictionary, with the candidate determinants as keys and the sums of the and (for PT2 correction only) contributions from all connected selected determinants as values. This allows for efficient perturbative selection in the next step, as well as for an estimate of the perturbative correction at the end of the selection process.
(ii) Perturbative refinement. The preselected determinants are then ranked by their first-order (Epstein-Nesbet) perturbative coefficient,
and only those with are chosen, where is by default equal to . The sum runs only over the variational determinants that survived the heat-bath criterion and is obtained virtually for free from the preselection step, as the sum of instantaneous contributions . The combination of both selection steps yields a compact determinant space in which CI can be performed while retaining the determinants that contribute most significantly to the energy.
(iii) CI Hamiltonian construction. The CI Hamiltonian in the selected determinant space is constructed by evaluating the matrix elements between the selected determinants. The connectivity of the selected determinants is determined similarly to Ref. 75, by introducing helper dictionaries of same--string and same--string determinants, which allow for efficient identification of determinants connected by same-spin excitations, and additionally a single--excitation dictionary, which stores for all existing -strings all -strings connected to them by a single excitation, allowing for efficient identification of determinants connected by opposite-spin excitations.
(iv) Diagonalization. The CI Hamiltonian in the selected determinant space is diagonalized directly for small spaces ( determinants) or with a Davidson solver for larger ones. An optional small-space prediagonalization provides a good initial guess, and several roots can be targeted simultaneously () for excited states; in the multi-state case the selection uses a state-maximum criterion, i.e. a determinant is added if its perturbative coefficient is significant for any of the targeted states, which avoids biasing the space toward the ground state.
The entire cycle, consisting of steps (i), (ii), (iii) and (iv), is repeated until either the determinant count reaches the specified target or the energy has converged to the requested tolerance.
Perturbative correction and variants. Once the variational space has converged, a second-order Epstein-Nesbet correction,
is evaluated, where runs over determinants outside the selected determinant space. This is computed by the same heat-bath machinery as the selection, but with a tighter threshold (and an adaptive variant ); a further instantaneous screening discards negligible contributions before they are fully assembled, which bounds the memory cost and also yields an uncertainty estimate of the correction. An uncontracted MP2 form and a renormalized variant, , are also available. CIPHI calculations can be stored to and restarted from a portable TREXIO dump – useful for tightening thresholds, computing the PT2 correction on a pre-converged space (a PT2-only mode), or storing the states of a multi-state calculation separately.
Like the rest of ElemCo.jl, the FCI and CIPHI solvers are formulated generically in the element type and do not assume Hermiticity, so both also operate for complex-valued and non-Hermitian Hamiltonians.
Molecular properties
Beyond energies, ElemCo.jl can evaluate one-electron molecular properties at the CD, DC, and CI levels. From the one-particle reduced density matrix (1-RDM), obtained by solving the associated CCSD/DCSD Lagrange-multiplier equations, the package can compute properties such as dipole moments. Molecular properties can be evaluated with (U)CCSD and (U)DCSD by setting @set cc properties=true. The corresponding properties are likewise available for FCI and CIPHI wavefunctions.
Furthermore, natural orbitals and their occupation numbers can be obtained by diagonalizing the 1-RDM. These natural orbitals and occupations can be written to a TREXIO dump for subsequent reuse, for example in active-space selection procedures.
Excited states
In ElemCo.jl, excited states, in addition to FCI or CIPHI (see above), can be computed using equation-of-motion CC (EOM-CC) theory,[78, 79] implemented at the EOM-CCSD level and its DC counterpart EOM-DCSD. Both methods are available in closed-shell, restricted open-shell (R), and spin-unrestricted (U) variants (EOM-RCCSD/EOM-UCCSD and EOM-RDCSD/EOM-UDCSD). The corresponding trial vectors and Hamiltonian matrices are formulated for general numeric types, making excited-state calculations for non-Hermitian and complex-valued Hamiltonians equally accessible.
To assess the accuracy of the EOM-DCSD implementation, we present a benchmark against CC3 for the QUEST3 benchmark set[80, 81] of vertical excitation energies. The results are shown in Figure 2. All EOM-DCSD results obtained in this work, together with the EOM-CCSD and CC3 benchmark data from QUEST3, are provided in the Supplementary Information, including the ordering of molecules and excited states used in the figure. The QUEST3 set comprises medium-sized molecules containing four to six non-hydrogen atoms. We consider 150 states with dominant single-excitation character for which both EOM-CCSD and CC3 reference data are available in the augmented triple-zeta basis.[82, 83] Consistent with previous studies,[84, 85, 86] EOM-DCSD outperforms EOM-CCSD for the vast majority of states. Of the 150 excitations considered, EOM-DCSD performs better for 133 states, both methods perform equally well for three states, while EOM-DCSD yields larger deviations from CC3 than EOM-CCSD for only 14 states. The mean absolute deviation (MAD) and root-mean-square deviation (RMSD), also shown in Figure 2, demonstrate that EOM-DCSD reduces the error (with respect to the CC3 reference) by more than a factor of two relative to EOM-CCSD. For EOM-CCSD, the largest deviation is observed for the state of glyoxal, amounting to 0.54 eV. For the same state, the EOM-DCSD deviation is only 0.24 eV, i.e., less than half that of EOM-CCSD. For EOM-DCSD, the largest deviation is observed for the state of pyrazine, with an error of 0.25 eV. For this state, the corresponding EOM-CCSD deviation is 0.41 eV, which is almost twice as large. It should also be noted that, in addition to being more accurate, EOM-DCSD is slightly less computationally expensive than EOM-CCSD. Like EOM-CCSD, EOM-DCSD formally scales as ; however, because certain terms in the residual equations are omitted, the computational prefactor is smaller.
Figure 2. Benchmark of 150 vertical single-excitation energies computed with EOM-DCSD and EOM-CCSD relative to CC3 for the QUEST3 benchmark set,[80, 81] which comprises medium-sized molecules containing four to six non-hydrogen atoms. All calculations employ the augmented triple-zeta basis. MAD, RMSD, and MAX denote the mean absolute deviation, root-mean-square deviation, and maximum deviation, respectively.
Non-Hermitian and Complex-Valued Hamiltonians
The majority of methods in ElemCo.jl do not assume Hermiticity of the Hamiltonian and can therefore be applied directly to non-Hermitian and to complex-valued Hamiltonians. This capability enables electronic-structure calculations based on integrals supplied through a generalized FCIDUMP file, making ElemCo.jl suitable for non-Hermitian theories such as transcorrelated methods.[19, 20, 21, 87] In particular, the xTC Hamiltonian (transcorrelated Hamiltonian excluding explicit three-body terms) can be generated with the (PY)TCHInt program,[88] and the resulting integrals can be read into ElemCo.jl for subsequent correlation calculations. The non-Hermiticity of the Hamiltonian is indicated by setting ST=1 (i.e., Similarity-Transformed) in the header of the FCIDUMP file.
Mixed electron-positron systems
ElemCo.jl can treat mixed electron–positron systems at the density-fitted Hartree–Fock (DF-HF) and second-order Møller–Plesset (MP2) levels within a restricted formalism. A positron is added to the system with @set wf npositron=1. Only a single positron is currently supported, so no positron–positron integrals arise.
For interfacing with external solvers, ElemCo.jl can write and read MOLDEN and FCIDUMP files for these systems. The MOLDEN files include an additional field indicating the presence of a positron and list the positron orbitals separately from the electronic ones; in the FCIDUMP files, a line of zeroes separates the electron–positron integrals from the purely electronic ones. These extensions make it possible to couple positron solvers with other codes, such as the variational and diffusion quantum Monte Carlo (VMC and DMC) methods of CASINO.
Interfaces and interoperability
ElemCo.jl is designed to interoperate with external quantum chemistry programs. A schematic overview of how ElemCo.jl can be used together with other programs and how data can be exchanged between them is shown in Figure 3.
Figure 3. Schematic overview of the input and output options available in ElemCo.jl, including supported interfaces and interoperability with external software.
For generating inputs in which the molecular geometry and basis set are specified and the corresponding integrals are evaluated via LIBCINT, JMol and MOLPRO provide interfaces, although their use is entirely optional. Integrals and Hamiltonians can be exchanged through the FCIDUMP format, which is routinely supported by many quantum chemistry programs. The FCIDUMP interface also supports the transfer of non-Hermitian Hamiltonians, including those used in transcorrelation and generated, for example, with TCHINT,[88] as well as model Hamiltonians generated with DUMPHAM.[89] The extended FCIDUMP format can also store the integral values in a binary NPY format using the ElemCo.jl’s storage convention, section II D, which is much more efficient for large systems than the standard ASCII format and can directly be read into ElemCo.jl.
Regarding the output, in addition to the standard text output summarizing the most important details and results of the quantum chemical calculations, all results are directly reusable in Julia. In particular, computed quantities can be stored as Julia objects and can therefore be processed further within user scripts. Wavefunction data such as orbitals and amplitudes can be written to portable dump files in the TREXIO format,[90] enabling interchange with other codes. Moreover, ElemCo.jl can generate MOLDEN files for orbital visualization, which can be viewed either directly in JMol or in other visualization programs.
Finally, ElemCo.jl exposes an interface to the ITensors.jl/ITensorMPS.jl libraries[91] for densitymatrix renormalization-group (DMRG)[92, 93] calculations.
Representative examples
In this section, we present a few complete ElemCo.jl input examples. Note that the examples are not meant to be exhaustive, but rather to illustrate the basic usage of ElemCo.jl and its macro interface. Besides, the program is under active development, and the input syntax or default settings may change in future releases. For a complete up-to-date list of available options and their default values, please refer to the ElemCo.jl documentation.[18]
DCSD from a geometry. In this example, a molecular geometry and a basis set are specified. ElemCo.jl first performs a density-fitted Hartree-Fock calculation and subsequently a DCSD calculation. Unless explicitly specified otherwise (here, the unit bohr is given explicitly), Bohr is used as the default unit for molecular geometries. Alternatively, geometries can be specified in Angstrom by writing angstrom instead of bohr, or by using the XYZ format (first line: number of atoms; second line: comment or empty line; subsequent lines: atom type and Cartesian coordinates in Angstrom). The basis set can be specified as a dictionary, as shown below, containing entries for the atomic orbital basis (AO), the HF auxiliary basis (JKFIT), and the correlation auxiliary basis (MPFIT). Alternatively, only the AO basis can be provided, for example through basis = "name_of_basis", in which case suitable JKFIT and MPFIT basis sets are selected automatically. For instance, specifying basis = "vdz" selects auxiliary basis sets at the valence-double-zeta level. The use of @print_input is optional.
using ElemCo
@print_input
geometry = "bohr"
O 0.000000 0.000000 -0.123995
H1 0.000000 1.431132 0.984106
H2 0.000000 -1.431132 0.984106"
basis = Dict("ao" => "cc-pVDZ",
"jkfit" => "cc-pvtz-jkfit",
"mpfit" => "cc-pvdz-mpfit")
@dfhf
@cc dcsd
CCSD(T) from an FCIDUMP file. In the following example, an FCIDUMP file named H2O.FCIDUMP is read and used as the input Hamiltonian for a CCSD(T) calculation. The energies from the CCSD(T) calculation are stored in the ordered dictionary ccsd_t_energies and can be directly accessed in further Julia code.
using ElemCo
fcidump = "H2O.FCIDUMP"
ccsd_t_energies = @cc ccsd(t)
EOM-DCSD with a local option block. In this example, a molecular geometry is read from the file "geometry.xyz". Subsequently, a HF calculation is performed, followed by an EOM-DCSD calculation. Options can be confined to a single method call so that they do not affect subsequent calculations. Here, the local option block is used to specify the number of EOM-DCSD states to be computed:
using ElemCo
@print_input
geometry = "geometry.xyz"
basis = "avtz"
@hf
@cc eom-dcsd begin
@set eom nstates=5
end
SVD-DC-CCSDT calculation. For the molecular geometry specified in "geometry.xyz" and a valence-double-zeta basis set, ElemCo.jl first performs a DF-HF calculation followed by an SVD-DC-CCSDT calculation. To assess convergence with respect to the size of the SVD basis, as controlled by the decomposition threshold , the option @set cc ampsvdtol can be used to tighten the truncation threshold. The default value is currently 1.e-5.
using ElemCo
geometry = "geometry.xyz"
basis = "vdz"
@dfhf
@cc svd-dc-ccsdt begin
@set cc ampsvdtol = 1.e-6
end
Scripting and invoking internal functions. In addition to the macro interface, ElemCo.jl allows users to call internal functions directly from the input file. In the example below, the doubles amplitudes T2 from a DCSD calculation are loaded together with the occupied-occupied-virtual-virtual integral block oovv. These quantities are then contracted to form the closed-shell doubles contribution to the correlation energy using the @mtensor macro, which follows Einstein summation conventions:
using ElemCo
using ElemCo.TensorTools # ints1, ints2, @mtensor
fcidump = "FCIDUMP"
@cc dcsd
T2 = @loadfile cc_amplitudes_2
oovv = ints2(EC, "oovv") # <ij|ab>, indexed [i,j,a,b]
@mtensor E = (2*T2[a,b,i,j] - T2[b,a,i,j])*oovv[i,j,a,b]
Any other quantity can be assembled in the same way: the required integral blocks can be obtained via ints1 and ints2, amplitudes can be loaded from disk, and the resulting expressions can be evaluated with @mtensor. This flexibility makes prototyping and testing new methods on top of ElemCo.jl straightforward.
Performance
In the following, the computational performance of ElemCo.jl is assessed using representative calculations on the phenol molecule consisting of 13 atoms and 50 electrons. The molecular geometry of phenol is provided in the Supplementary Material. The calculations employ an augmented triple-zeta basis (aug-cc-pVTZ) [82, 83] comprising 460 atomic orbitals, 1003 auxiliary functions in the JKFIT basis and 1017 auxiliary functions in the MPFIT basis. The DF-HF energy is -305.670426 hartree, and the corresponding running time is less than 10 seconds, making its cost negligible compared to that of the subsequent correlated calculations (apart from DF-MP2, which is even faster than DF-HF).
With the frozen core approximation, the correlation calculations involve 18 occupied orbitals and 435 virtual orbitals. Table III summarizes the timings and energies obtained with different electron-correlation methods. As expected, MP2 is by far the fastest method. DCSD and CCSD are comparable in the overall cost, with DCSD being slightly faster as it omits some computationally expensive exchange diagrams, as discussed in Section III D. As expected, the DCSD energy is closer to the higher-level CCSD(T) and SVD-DC-CCSDT values, as has been consistently observed for this method.[22, 23, 64]
| Method | , s | , s | , | |
| DF-MP2 | 2 | - | - | -1.205019 |
| DCSD | 1456 | 13 | 112 | -1.259989 |
| CCSD | 1526 | 12 | 127 | -1.223712 |
| CCSD(T) | 5768 | - | - | -1.287759 |
| SVD-DC-CCSDT | 6344 | 12 | 346 | -1.284869 |
Table III. Timings and correlation energies obtained with different electron correlation methods for phenol (aug-cc-pVTZ, 36 correlated electrons in 453 orbitals). Here, denotes the total wall time required by the respective correlation method, while is the average time per iteration. The time given for CCSD(T) corresponds only to the evaluation of the (T) correction. The number of iterations is given by , and denotes the correlation energy. All calculations were performed on 12-core Intel Xeon Gold 6128 nodes with a rate of 3.40 GHz.
For the present system, the perturbative (T) is approximately four times more expensive than DCSD or CCSD. Despite containing iterative triples amplitudes, with the wall time of the SVD-DC-CCSDT calculation is of the same order of magnitude as the non-iterative (T) correction. This efficiency is a consequence of the tensor-decomposed approach described in Section III F, which substantially reduces the computational scaling and the prefactor compared to conventional DC-CCSDT.
Illustrative application
In this prototypical application, we demonstrate how the FCIDUMP interface to ElemCo.jl can be used to study complex systems at a high level of quantum-chemical theory. Specifically, we address the formation energy of a negatively charged vacancy in phosphorene, with particular focus on the post-CCSD(T) contribution, as a follow-up to the CCSD(T) benchmark of [94].
Phosphorene is a single two-dimensional layer that stacks to form bulk black phosphorus. A highly accurate quantum-chemical description of black phosphorus is of particular interest in the context of a recent controversy concerning discrepancies between CCSD(T) and quantum Monte Carlo (QMC) calculations of intermolecular interactions in large dimers.[95, 96, 97, 98] Interestingly, a possibly related effect has also been observed for the exfoliation energy of black phosphorus: despite overall good agreement between the two approaches, CCSD(T) predicts[99] a slightly stronger binding than QMC.[100] Here, we focus not on exfoliation but rather on the formation energy of a negatively charged vacancy. Nevertheless, the results of [94] demonstrate that van der Waals dispersion makes a sizable contribution to the vacancy formation energy, suggesting that an estimate of the post-CCSD(T) correction provides important additional information.
We calculate the defect formation energy using the recently developed aperiodic defect model (ADM)[101] as implemented in the CRYSCOR code.[102] In this model, a fragment containing the defect is embedded in the periodic Hartree-Fock mean field of the pristine crystal. The fragment-environment partitioning is based on the localized orbitals of the pristine structure and does not require cutting out a physical fragment. The ADM offers several key advantages over conventional supercell calculations. In particular, it does not periodically repeat the defect image, as in the supercell approach, and thus avoids unphysical interactions between defect images. This is especially important for charged defects such as the one considered here. Another important feature of the ADM is that the embedding field is included in the fragment’s one-electron Hamiltonian, effectively reducing the problem to a molecule-like one. The matrix elements of the resulting effective fragment Hamiltonian in the basis of the fragment HF orbitals can then be written in FCIDUMP format and passed to ElemCo.jl. A detailed description of the ADM and its formalism can be found in [101].
In this work, we calculate the correlation contribution to the formation energy as the energy of the defective fragment plus the energy of an isolated phosphorus atom in its quartet ground state, minus the energy of the corresponding fragment in the pristine structure; see Eq. (6) of [94]. We employ an 18-atom fragment from the defective structure, shown in Figure 4, which corresponds to a 19-atom fragment for the pristine structure. The structural and computational parameters for the ADM calculations and the generation of the FCIDUMP files are the same as in [94], including the POB-TZVP-rev2 basis set.[103] The post-HF methods, up to SVD-DC-CCSDT, are run with ElemCo.jl using the FCIDUMP-based input shown in § 1.4. With the frozen-core approximation employed, the 18-atom fragment of the defective structure contains 76 electrons in 306 orbitals, while the 19-atom fragment of the pristine structure contains 80 electrons in 323 orbitals. The electron counts are smaller than the corresponding numbers of atomic valence electrons because some of the bonding electrons of the fragment-edge atoms belong to localized orbitals assigned to the environment and are therefore excluded from the fragment.[94]
Figure 4. The 18-atom fragment containing the negatively charged vacancy in phosphorene. Atoms belonging to the fragment used in the ADM are shown in pink, while atoms of the embedding environment are shown in gold. For details of the geometry optimization, see Ref. 94.
For the phosphorus atom, we perform the complete calculation on the ElemCo.jl side using the same orbital and auxiliary basis sets as in the ADM calculations. The atom is surrounded by 11 ghost atoms taken from the pristine structure to account for basis-set superposition error. The quartet state is treated within the unrestricted formalism. For UDC-CCSDT, the SVD approximation is not applied because it is not implemented for the open-shell case.
The calculated formation energies, or more precisely their correlation-energy contributions, are listed in Table IV. First, we note that the post-CCSD(T) correction is rather small, both in absolute terms ( eV or 0.25 kcal/mol) and relative to the magnitude of the formation energy itself. Furthermore, it is smaller than the estimated remaining uncertainty in the benchmark of Ref. 94 due to finite-size effects and basis-set incompleteness. This suggests that the CCSD(T) level is already sufficient for the required accuracy of for the phosphorene vacancy.
| Method | , eV | , h | ||
| MP2 | 1.061 | 0.0 | ||
| CCSD | 0.865 | 0.5 | ||
| CCSD(T) | 0.930 | 4.1 | ||
| SVD treatment | ||||
| , eV | , h | , eV | , h | |
| SVD-CCSD(T) | 0.915 | 1.7 | 0.929 | 4.3 |
| 0.015 | 0.001 | |||
| SVD-DC-CCSDT | 0.904 | 6.7 | 0.918 | 88.9 |
| -0.011 | -0.011 |
Table IV. Correlation-energy contribution to the defect formation energy at different quantum-chemical levels and the corresponding timings (total wall time) on a 44-CPU node. For the SVD-based calculations, decomposition thresholds and were employed. The SVD error is evaluated at the (T) level as . The post-CCSD(T) formation-energy contribution is defined as . The timings correspond to the calculations on the 19-atom fragment for the pristine structure (80 electrons in 323 orbitals) on a 12-core Intel Xeon Gold 6128 node with a rate of 3.40 GHz.
As this is an illustrative application, we have not attempted to converge with respect to fragment size and/or basis-set size. Nevertheless, its small magnitude already for the 18/19-atom fragments and the POB-VTZ-rev2 basis suggests that the converged value is unlikely to become significant.
As for the SVD approximation, with a decomposition threshold of , the SVD error in becomes essentially negligible. With , the error in is still appreciably small, although not as small as for . Importantly, the post-CCSD(T) correction looks fully converged already at . Remarkably, the computational time of SVD-DC-CCSDT with is comparable to that of CCSD(T).
Conclusions and outlook
ElemCo.jl is a modern quantum chemistry package written in Julia for molecular electron-correlation calculations. It combines an intuitive, macro-driven user interface that facilitates both routine calculations and rapid method prototyping with a transparent and extensible code base. State-of-the-art infrastructure for solvers, tensor-decomposition techniques like DF or SVD, wavefunction-data management, and interoperability with external software makes it a versatile platform for electronic-structure method development. At the same time, Julia enables high computational performance without sacrificing readability or ease of development. For input generation and visualization of results, the companion program JMol is available.
As a result, ElemCo.jl provides a flexible platform for both electronic-structure method developers and practitioners applying quantum-chemical methods. ElemCo.jl is free and open-source software; its source code, documentation, and examples are available online at https://github.com/fkfest/ElemCo.jl and https://elem.co.il. The package offers a broad range of methods, spanning HF and MP2 to CC, DC, CI, and EQM approaches in closed-shell, restricted-open-shell, and unrestricted formulations.
Importantly, several methods are made openly available through ElemCo.jl for the first time or are, to the best of our knowledge, not freely accessible in any other package. These include tensor-decomposed DC methods such as SVD-DC-CCSDT, two-determinant and fixed-reference DC models, and the CIPHI variant of the selected-CI method. Furthermore, support for non-Hermitian Hamiltonians across many methods makes ElemCo.jl a natural platform for transcorrelated electronic-structure theory.
Additionally, we demonstrated the superior accuracy of EOM-DCSD compared to EOM-CCSD. Both methods are available in closed-shell, restricted open-shell, and spin-unrestricted variants.
Development of ElemCo.jl is ongoing and encompasses new method development, performance optimization, expansion of the package’s capabilities, and additional interfaces to external software. Current efforts include the completion of SVD-DCSD and related tensor-decomposed methods, quasi-variational approaches, extension of various methods to include positron affinity calculations, improving MCSCF and CASSCF, parallelization, and the implementation of additional molecular properties and analytical gradients, and continued improvements in computational performance.
Supplementary Material
The Supplementary Material contains the EOM-DCSD benchmark data discussed in Section III J, the geometry of phenol used in the performance calculations presented in § 1.5, as well as the raw data and orbital-space sizes for the vacancy formation-energy calculations in Section VII.
Acknowledgments
Financial support from the Max Planck Society is gratefully acknowledged. C.R. gratefully acknowledges Studienstiftung des deutschen Volkes for a Master’s scholarship and Tiemann Stiftung for financial support of research visits to the Max Planck Institute for Solid State Research.
Author Declarations
Conflict of Interest
The authors have no conflicts to disclose.
Data Availability
The ElemCo.jl source code and documentation are openly available at https://github.com/fkfest/ElemCo.jl and https://elem.co.il.
References
- [1]G. D. Purvis and R. J. Bartlett, “A full coupled-cluster singles and doubles model: The inclusion of disconnected triples,” J. Chem. Phys. 76, 1910–1918 (1982).
- [2]K. Raghavachari, G. W. Trucks, J. A. Pople, and M. Head-Gordon, “A fifth-order perturbation comparison of electron correlation theories,” Chem. Phys. Lett. 157, 479–483 (1989).
- [3]R. J. Bartlett and M. Musiał, “Coupled-cluster theory in quantum chemistry,” Rev. Mod. Phys. 79, 291–352 (2007).
- [4]H.-J. Werner, P. J. Knowles, G. Knizia, F. R. Manby, and M. Schütz, “Molpro: A general-purpose quantum chemistry program package,” WIREs Comput. Mol. Sci. 2, 242–253 (2012).DOI
- [5]H.-J. Werner, P. J. Knowles, F. R. Manby, J. A. Black, K. Doll, A. Heßelmann, D. Kats, A. Köhn, T. Korona, D. A. Kreplin, et al., “The molpro quantum chemistry package,” J. Chem. Phys. 152, 144107 (2020).
- [6]A. D. Matthews, L. Cheng, M. E. Harding, F. Lipparini, S. Stopkowicz, T.-C. Jagau, P. G. Szalay, J. Gauss, and J. F. Stanton, “Coupled-cluster techniques for computational chemistry: The CFOUR program package,” J. Chem. Phys. 152, 214108 (2020).DOI
- [7]F. Neese, F. Wennmohs, U. Becker, and C. Riplinger, “The ORCA quantum chemistry program package,” J. Chem. Phys. 152, 224108 (2020).
- [8]F. Furche, R. Ahlrichs, C. Hättig, W. Klopper, M. Sierka, and F. Weigend, “Turbomole,” Wiley Interdiscip. Rev. Comput. Mol. Sci. 4, 91–100 (2014).
- [9]M. Kállay, P. R. Nagy, D. Mester, Z. Rolik, G. Samu, J. Csontos, J. Csóka, P. B. Szabó, L. Gyevi-Nagy, B. Hégely, et al., “The MRCC program system: Accurate quantum chemistry from water to proteins,” J. Chem. Phys. 152, 074107 (2020).
- [10]E. Aprà et al., “NWChem: Past, present, and future,” J. Chem. Phys. 152, 184102 (2020).
- [11]G. M. J. Barca et al., “Recent developments in the general atomic and molecular electronic structure system,” J. Chem. Phys. 152, 154102 (2020).DOI
- [12]Q. Sun, T. C. Berkelbach, N. S. Blunt, G. H. Booth, S. Guo, Z. Li, J. Liu, J. D. McClain, E. R. Sayfutyarova, S. Sharma, S. Wouters, and G. K.-L. Chan, “PySCF: the Python-based simulations of chemistry framework,” WIREs Comput. Mol. Sci. 8, e1340 (2018).arxiv.org/abs/1701.08223
- [13]Q. Sun, X. Zhang, S. Banerjee, P. Bao, M. Barbry, N. S. Blunt, et al., “Recent developments in the PySCF program package,” J. Chem. Phys. 153, 024109 (2020).DOI
- [14]K. Boguslawski, F. Brzek, R. Chakraborty, K. Cieslak, S. Jahani, A. Leszczyk, A. Nowak, E. Sujkowski, J. Swierczynski, S. Ahmadhkani, D. Kedziera, M. H. Kriebel, P. S. Zuchowski, and P. Temmer, “Pybest: Improved functionality and enhanced performance,” Comput. Phys. Commun. 297, 109049 (2024).
- [15]J. Bezanson, A. Edelman, S. Karpinski, and V. B. Shah, “Julia: A fresh approach to numerical computing,” SIAM Rev. 59, 65–98 (2017).
- [16]M. F. Herbst, A. Levitt, and E. Cancès, “Dftk: A julian approach for simulating electrons in solids,” Proc. JuliaCon Conf. 3, 69 (2021).DOI
- [17]G. J. R. Aroeira, M. M. Davis, J. M. Turney, and H. F. I. Schaefer, “Fermi.jl: A Modern Design for Quantum Chemistry,” J. Chem. Theory Comput. 18, 677 (2022).DOI
- [18]D. Kats, C. Rickert, T. Schravogel, K. Simula, D. Usvyat, and F. Wu, “ElemCo.jl: Julia program package for electron correlation methods,” (2026), https://github.com/fkfest/ElemCo.jl-devel; https://elem.co.il.arxiv.org/abs/2608.27177
- [19]S. F. Boys and N. C. Handy, “A calculation for the energies and wavefunctions for states of neon with full electronic correlation accuracy,” Proc. R. Soc. of Lond. A 310, 63–78 (1969).DOI
- [20]T. Schraivogel, A. J. Cohen, A. Alavi, and D. Kats, “Transcorrelated coupled cluster methods,” J. Chem. Phys. 155, 191101 (2021).
- [21]T. Schraivogel and D. Kats, “Transcorrelated coupled cluster methods. II. molecular systems,” J. Chem. Phys. 158, 214106 (2023).DOI
- [22]D. Kats and F. R. Manby, “Communication: The distinguishable cluster approximation,” J. Chem. Phys. 139, 021102 (2013).
- [23]D. Kats, “Communication: The distinguishable cluster approximation. II. The role of orbital relaxation,” J. Chem. Phys. 141, 061101 (2014).DOI
- [24]J. L. Whitten, “Coulombic potential energy integrals and approximations,” J. Chem. Phys. 58, 4496–4501 (1973).
- [25]E. J. Baerends, D. Ellis, and P. Ros, “Self-consistent molecular hartree–fock–slater calculations i. the computational procedure,” Chem. Phys. 2, 41–51 (1973).
- [26]B. I. Dunlap, J. W. Connolly, and J. Sabin, “Some approximations in applications of x. cap alpha. theory,” J. Chem. Phys. 71, 4993–4999 (1979).
- [27]O. Vahtras, J. Almlöf, and M. W. Feyereisen, “Integral approximations for LCAO-SCF calculations,” Chem. Phys. Lett. 213, 514–518 (1993).DOI
- [28]F. Weigend, “A fully direct RI-HF algorithm: Implementation, optimised auxiliary basis sets, demonstration of accuracy and efficiency,” Phys. Chem. Chem. Phys. 4, 4285–4291 (2002).DOI
- [29]N. H. Beebe and J. Linderberg, “Simplifications in the generation and transformation of two-electron integrals in molecular calculations,” Int. J. Quantum Chem. 12, 683–705 (1977).DOI
- [30]H. Koch, A. Sánchez de Merás, and T. B. Pedersen, “Reduced scaling in electronic structure calculations using Cholesky decompositions,” J. Chem. Phys. 118, 9481–9484 (2003).DOI
- [31]S. D. Folkestad, E. F. Kjønstad, and H. Koch, “An efficient algorithm for Cholesky decomposition of electron repulsion integrals,” J. Chem. Phys. 150, 194112 (2019).
- [32]J. Almlöf, “Elimination of energy denominators in møller–plesset perturbation theory by a laplace transform approach,” Chem. Phys. Lett. 181, 319–320 (1991).
- [33]M. Häser and J. Almlöf, “Laplace transform techniques in mo/ller–plesset perturbation theory,” J. Chem. Phys. 96, 489–494 (1992).DOI
- [34]D. Kats, D. Usvyat, and M. Schütz, “On the use of the laplace transform in local correlation methods,” Phys. Chem. Chem. Phys. 10, 3430–3439 (2008).DOI
- [35]T. Kinoshita, O. Hino, and R. J. Bartlett, “Singular value decomposition approach for the approximate coupled-cluster method,” J. Chem. Phys. 119, 7756–7762 (2003).DOI
- [36]G. Knizia, “Intrinsic atomic orbitals: An unbiased bridge between quantum theory and chemical concepts,” J. Chem. Theory Comput. 9, 4834–4843 (2013).DOI
- [37]J. Pipek and P. G. Mezey, “A fast intrinsic localization procedure applicable for ab initio and semiempirical linear combination of atomic orbital wave functions,” J. Chem. Phys. 90, 4916–4926 (1989).DOI
- [38]J. M. Foster and S. F. Boys, “Construction of some molecular orbitals to be approximately invariant for changes from one molecule to another,” Rev. Mod. Phys. 32, 300–302 (1960).DOI
- [39]P. Pulay and S. Saebø, “Orbital-invariant formulation and second-order gradient evaluation in møller–plesset perturbation theory,” Theor. Chim. Acta 69, 357–368 (1986).DOI
- [40]S. Saebo and P. Pulay, “Local treatment of electron correlation,” Annu. Rev. Phys. Chem. 44, 213–236 (1993).
- [41]I. Shavitt, C. Bender, A. Pipano, and R. Hosteny, “The iterative calculation of several of the lowest or highest eigenvalues and corresponding eigenvectors of very large symmetric matrices,” J. Chem. Phys. 11, 90–108 (1973).DOI
- [42]P. Pulay, “Convergence acceleration of iterative sequences. the case of scf iteration,” Chem. Phys. Lett. 73, 393–398 (1980).DOI
- [43]T. P. Hamilton and P. Pulay, “Direct inversion in the iterative subspace (diis) optimization of open-shell, excited-state, and small multiconfiguration scf wave functions,” J. Chem. Phys. 84, 5728–5734 (1986).DOI
- [44]C. Rickert, D. Usvyat, and D. Kats, “Tensor decomposed distinguishable cluster. I. Triples decomposition,” J. Chem. Phys. 163, 064103 (2025).
- [45]S. Lambie, C. Rickert, D. Usvyat, A. Alavi, and D. Kats, “Benchmarking distinguishable cluster methods to platinum standard ccsdt (q) non-covalent interaction energies in the a24 dataset,” J. Chem. Phys. 163, 111101 (2025).
- [46]P. G. Szalay and R. J. Bartlett, “Analytic energy gradients for the two-determinant coupled cluster method with application to singlet excited states of butadiene and ozone,” J. Chem. Phys. 101, 4936 (1994).DOI
- [47]T. Schraivogel and D. Kats, “Two determinant distinguishable cluster,” J. Chem. Phys. 160, 124109 (2024).
- [48]T. Schraivogel and D. Kats, “Accuracy of the distinguishable cluster approximation for triple excitations for open-shell molecules and excited states,” J. Chem. Phys. 155, 064101 (2021).DOI
- [49]Q. Sun, “Libcint: An efficient general integral library for gaussian basis functions,” J. Comput. Chem. 36, 1664 (2015).
- [50]ElemCo.jl developers, “JLmol: input generator and molecular orbital visualizer for ElemCo.jl,” (2026), https://github.com/kflfe st/jlmol/; web version: https://app.jmol.com.
- [51]“Jmol: an open-source java viewer for chemical structures in 3d,” https://jmol.org
- [52]“JSmol: an open-source html5 viewer for chemical structures in 3d,” http://wiki.jmol.org/index.php/JSmol.
- [53]P. J. Knowles and N. C. Handy, “A determinant based full configuration interaction program,” Comput. Phys. Commun. 54, 75 (1989).
- [54]L. Devos, M. Van Damme, J. Haegeman, and contributors, “TensorOperations.jl,” (2025), github.com/QuantumKitHub/TensorOperations.jl.
- [55]B. P. Pritchard, D. Altarawly, B. Didier, T. D. Gibson, and T. L. Windus, “New Basis Set Exchange: An Open, Up-to-Date Resource for the Molecular Sciences Community,” J. Chem. Inf. Model. 59, 4814 (2019).DOI
- [56]D. A. Kreplin, P. J. Knowles, and H.-J. Werner, “MCSCF optimization revisited. II. Combined first- and second-order orbital optimization for large molecules,” J. Chem. Phys. 152, 074102 (2020).DOI
- [57]U. Meier and V. Staemmler, “An efficient first-order casscf method based on the renormalized fock-operator technique,” Theor. Chem. Acc. 76, 95–111 (1989).DOI
- [58]D. R. Yarkony, “Comment on the use of the augmented matrix in mcscf theory,” Chem. Phys. Lett. 77, 634–635 (1981).DOI
- [59]S. A. Kucharski and R. J. Bartlett, “Noniterative energy corrections through fifth-order to the coupled cluster singles and doubles method,” J. Chem. Phys. 108, 5243 (1998).DOI
- [60]P. J. Knowles, C. Hampel, and H.-J. Werner, “Coupled cluster theory for high spin, open shell reference wave functions,” J. Chem. Phys. 99, 5219 (1993).
- [61]P. J. Knowles, C. Hampel, and H.-J. Werner, “Erratum: “Coupled cluster theory for high spin, open shell reference wave functions”,” J. Chem. Phys. 112, 3106 (2000).
- [62]D. Kats and T. Schraivogel, “Quantwo: second-quantization program,” (2024), github.com/kflest/quantwo.
- [64]D. Kats, D. Kreplin, H.-J. Werner, and F. R. Manby, “Accurate thermochemistry from explicitly correlated distinguishable cluster approximation,” J. Chem. Phys. 142, 064111 (2015).DOI
- [65]D. Kats, “The distinguishable cluster approach from a screened Coulomb formalism,” J. Chem. Phys. 149, 191101 (2018).
- [66]D. Kats, “Improving the distinguishable cluster results: spin-component scaling,” Mol. Phys. 116, 1435 (2018).DOI
- [67]D. Kats, D. Usvyat, and F. R. Manby, “Particle–hole symmetry in many-body theories of electron correlation,” Mol. Phys. 116, 1496 (2018).DOI
- [68]D. Kats and A. Köhn, “On the distinguishable cluster approximation for triple excitations,” J. Chem. Phys. 150, 151101 (2019).
- [69]V. Rishi and E. F. Valeev, “Can the distinguishable cluster approximation be improved systematically by including connected triples?” J. Chem. Phys. 151, 064102 (2019).DOI
- [70]J. B. Robinson and P. J. Knowles, “Quasi-variational coupled cluster theory,” The Journal of Chemical Physics 136, 054114 (2012).DOI
- [71]J. A. Black and P. J. Knowles, “Statistical analysis of activation and reaction energies with quasi-variational coupled-cluster theory,” Mol. Phys. 116, 1421 (2018).DOI
- [72]F. Wu and D. Kats, “Quasi-variational distinguishable cluster theory,” (2026), submitted to J. Chem. Phys.
- [73]B. Huron, J. P. Malrieu, and P. Rancurel, “Iterative perturbation calculations of ground and excited state energies from multiconfigurational zeroth-order wavefunctions,” J. Chem. Phys. 58, 5745–5759 (1973).DOI
- [74]A. A. Holmes, N. M. Tubman, and C. J. Umrigar, “Heat-Bath Configuration Interaction: An Efficient Selected Configuration Interaction Algorithm Inspired by Heat-Bath Sampling,” J. Chem. Theory Comput. 12, 3674–3680 (2016).DOI
- [75]S. Sharma, A. A. Holmes, G. Jeanmairet, A. Alavi, and C. J. Umrigar, “Semistochastic Heat-Bath Configuration Interaction Method: Selected Configuration Interaction with Semistochastic Perturbation Theory,” J. Chem. Theory Comput. 13, 1595–1604 (2017).
- [76]M. Ugandi and M. Roemelt, “A configuration-based heatbath-ci for spin-adapted multireference electronic structure calculations with large active spaces,” J. Comput. Chem. 44, 2374 (2023).DOI
- [77]N. M. Tubman, C. D. Freeman, D. S. Levine, D. Haut, M. Head-Gordon, and K. B. Whaley, “Modern approaches to exact diagonalization and selected configuration interaction with the adaptive sampling ci method,” J. Chem. Theory Comput. 16, 2139 (2020).DOI
- [78]J. F. Stanton and R. J. Bartlett, “The equation of motion coupled-cluster method. a systematic biorthogonal approach to molecular excitation energies, transition probabilities, and excited state properties,” J. Chem. Phys. 98, 7029–7039 (1993).DOI
- [79]H. Koch and P. Jørgensen, “Coupled cluster response functions,” J. Chem. Phys. 93, 3333–3344 (1990).
- [80]P.-F. Loos, F. Lipparini, M. Boggio-Pasqua, A. Scemama, and D. Jacquemin, “A mountaineering strategy to excited states: Highly accurate energies and benchmarks for medium sized molecules,” J. Chem. Theory Comput. 16, 1711–1741 (2020).DOI
- [81]M. Véril, A. Scemama, M. Caffarel, F. Lipparini, M. Boggio-Pasqua, D. Jacquemin, and P.-F. Loos, “Questdb: A database of highly accurate excitation energies for the electronic structure community,” Wiley Interdiscip. Rev. Comput. Mol. Sci. 11, e1517 (2021).DOI
- [82]T. H. Dunning Jr, “Gaussian basis sets for use in correlated molecular calculations. i. the atoms boron through neon and hydrogen,” J. Chem. Phys. 90, 1007–1023 (1989).DOI
- [83]F. Weigend, A. Köhn, and C. Hätty, “Efficient use of the correlation consistent basis sets in resolution of the identity mp2 calculations,” J. Chem. Phys. 116, 3175–3183 (2002).DOI
- [84]V. Rishi, A. Perera, and R. J. Bartlett, “Assessing the distinguishable cluster approximation based on the triple bond-breaking in the nitrogen molecule,” J. Chem. Phys. 144, 124117 (2016).DOI
- [85]V. Rishi, A. Perera, M. Nooijen, and R. J. Bartlett, “Excited states from modified coupled cluster methods: Are they any better than eom ccsd?” J. Chem. Phys. 146, 144104 (2017).DOI
- [86]V. Rishi, M. Ravi, A. Perera, and R. J. Bartlett, “Dark doubly excited states with modified coupled cluster models: A reliable compromise between cost and accuracy?” J. Phys. Chem. A 127, 828–834 (2023).
- [87]E. M. Christlmaier, T. Schraivogel, P. López Ríos, A. Alavi, and D. Kats, “xTC: An efficient treatment of three-body interactions in transcorrelated methods,” J. Chem. Phys. 159, 041113 (2023).
- [88]J. P. Haupt, P. López Ríos, E. M. C. Christlmaier, M.-A. Filip, K. Simula, J. Hauskrecht, Y. Cheng, K. Liao, W. Dobrautz, K. Guther, A. J. Cohen, D. Kats, and A. Alavi, “TCHInt: Transcorrelated hamiltonian integrals library,”
- [89]D. Kats and T. Schraivogel, “DumpHam: Fcidump manipulation program,” github.com/fkfest/dumpham.
- [90]E. Posenitsky, V. G. Chilkuri, A. Ammar, M. Hapka, K. Pernal, R. Shinde, E. J. Landinez Borda, C. Filippi, K. Nakano, O. Kohulák, et al., “TREXIO: A file format and library for quantum chemistry,” J. Chem. Phys. 158, 174801 (2023).DOI
- [91]M. Fishman, S. R. White, and E. M. Stoudenmire, “The Tensor network software library for tensor network calculations,” SciPost Phys. Codebases , 4 (2022).DOI
- [92]S. R. White, “Density matrix formulation for quantum renormalization groups,” Phys. Rev. Lett. 69, 2863–2866 (1992).
- [93]G. K.-L. Chan and S. Sharma, “The density matrix renormalization group in quantum chemistry,” Annu. Rev. Phys. Chem. 62, 465–481 (2011).
- [94]C. Rickert, L. Barta, E.-C. Flach, D. Kats, and D. Usvyat, “Application of the aperiodic defect model to a negatively charged monovacancy in phosphorene,” J. Chem. Phys. 164, 234118 (2026).
- [95]Y. S. Al-Hamdani, P. R. Nagy, A. Zen, D. Barton, M. Kállay, J. G. Brandenburg, A. Tkatchenko, “Interactions between large molecules pose a puzzle for reference quantum mechanical methods,” Nat. Comm. 12, 3297 (2021).
- [96]S. Lambie, D. Kats, D. Usvyat, A. Alavi, “On the applicability of CCSD(T) for dispersion interactions in large conjugated systems,” J. Chem. Phys 162, 114112 (2025).
- [97]S. Lambie, P. L. Rios, D. Kats, and A. Alavi, “Nodal error behind discrepancies between coupled cluster and diffusion monte carlo,” J. Chem. Theory Comput. 22, 7693 (2026).DOI
- [98]C. Schäfer, A. Irmler, A. Gallo, and A. Grünewis, “Understanding discrepancies in noncovalent interaction energies from wave-function theories for large molecules,” Nat. Commun. 16, 9108 (2025).
- [99]M. Schütz, L. Maschio, A. J. Karttunen, and D. Usvyat, “Exfoliation energy of black phosphorus revisited: A coupled cluster benchmark,” J. Phys. Chem. Lett. 8, 1290 (2017).DOI
- [100]J. Shulenburger, A. Bazewski, Z. Zhu, J. Guan, and D. Tománek, “The nature of the interlayer interaction in bulk and few-layer phosphorus,” Nano Lett. 15, 8170 (2015).DOI
- [101]R. H. Lawler, D. Kats, L. Maschio, N. A. Bogdanov, A. Alavi, A. N. Alexandrovа, and D. Usvyat, “Aperiodic defects in periodic solids,” J. Chem. Phys. 163, 084105 (2025).
- [102]D. Usvyat, L. Maschio, and M. Schütz, “Periodic and fragment models based on the local correlation approach,” WIREs: Comput. Mol. Sci. 8, e1357 (2018).
- [103]D. Vilela Oliveira, J. Laun, M. F. Peintinger, and T. Bredow, “Bsse-correction scheme for consistent gaussian basis sets of double- and triple-zeta valence with polarization quality for solid-state calculations,” J. Comput. Chem. 40, 2364 (2019).DOI