Analytic energy gradients for the coupled-cluster singles and doubles with perturbative triples method with the density-fitting approximation

Bozkaya, Uğur; Sherrill, C. David

doi:10.1063/1.4994918

An efficient implementation of analytic gradients for the coupled-cluster singles and doubles with perturbative triples [CCSD(T)] method with the density-fitting (DF) approximation, denoted as DF-CCSD(T), is reported. For the molecules considered, the DF approach substantially accelerates conventional CCSD(T) analytic gradients due to the reduced input/output time and the acceleration of the so-called “gradient terms”: formation of particle density matrices (PDMs), computation of the generalized Fock-matrix (GFM), solution of the Z-vector equation, formation of the effective PDMs and GFM, back-transformation of the PDMs and GFM, from the molecular orbital to the atomic orbital (AO) basis, and computation of gradients in the AO basis. For the largest member of the molecular test set considered (⁠ $C_{6} H_{14}$ ⁠), the computational times for analytic gradients (with the correlation-consistent polarized valence triple- $ζ$ basis set in serial) are 106.2 [CCSD(T)] and 49.8 [DF-CCSD(T)] h, a speedup of more than 2-fold. In the evaluation of gradient terms, the DF approach completely avoids the use of four-index two-electron integrals. Similar to our previous studies on DF-second-order Møller–Plesset perturbation theory and DF-CCSD gradients, our formalism employs 2- and 3-index two-particle density matrices (TPDMs) instead of 4-index TPDMs. Errors introduced by the DF approximation are negligible for equilibrium geometries and harmonic vibrational frequencies.

I. INTRODUCTION

It has been demonstrated that coupled-cluster (CC) methods are accurate for the prediction of molecular properties.^1–6 The coupled-cluster singles and doubles (CCSD) method⁷ provides quite accurate results for most molecular systems at equilibrium geometries, but nevertheless a triple excitation correction is required to obtain high accuracy.^8–14 The coupled-cluster singles and doubles with perturbative triples [CCSD(T)] method^11,12,15 provides excellent results for a broad range of chemical systems near equilibrium geometries.^13,16–25

Analytic gradient methods are very helpful to locate stationary points on potential energy surfaces, as well as for the computation of one-electron properties.^26–30 Following the Z-vector method of Handy and Schaefer,³¹ Adamowicz et al.³² presented the initial algebraic expressions for analytic gradients of the CCSD method. In 1985, the analytic gradient for the coupled-cluster doubles (CCD) method was implemented by Fitzgerald et al.³³ In a 1987 study, Scheiner et al.³⁴ implemented the analytic gradient for the closed-shell CCSD method. Later, several CCSD analytic first derivative papers were also published.^35–39 The first implementation of CCSD(T) analytic gradients was reported by Scuseria for the closed-shell restricted Hartree-Fock (RHF) reference,⁴⁰ later improved by Lee and Rendell.⁴¹ In 1992 and 1993 studies, the CCSD(T) analytic gradients for unrestricted and restricted open-shell Hartree-Fock (UHF and ROHF) references were reported by Watts et al.^22,42 In 2003, Hald et al.⁴³ reported an integral direct implementation of CCSD(T) analytic gradients.

Approximate factorization of the electron repulsion integrals (ERIs) has been of considerable interest in modern quantum chemistry.^44–65 The most common ERI decomposition technique is density fitting (DF).^{44–51,58–65} In the DF approximation, the four-index ERIs are expanded in terms of three-index tensors. Another popular integral tensor decomposition approach is Cholesky decomposition (CD).^{54–59,62,63} The DF and CD approximations are very helpful in reducing the cost of integral transformations and the storage requirements of the ERI tensor. In the DF approximation, integrals of the primary basis set are expanded in terms of a pre-defined auxiliary basis set, whereas in the CD approach, the three-index factors are generated from the primary basis set.

Analytic gradients for the density-fitted post-Hartree–Fock (HF) methods have been reported for second-order Møller–Plesset perturbation theory (MP2) with conventional HF reference (RI-MP2),^66–68 as well as with the DF-HF reference,⁶⁰ coupled cluster model CC2,^69,70 local MP2 (DF-LMP2),⁷¹ complete active space second-order perturbation theory (DF-CASPT2),⁷² orbital-optimized MP2 (DF-OMP2),⁶¹ local time-dependent coupled cluster response theory (DF-TD-LCC2),⁷³ third-order Møller–Plesset perturbation theory (DF-MP3),⁷⁴ the DF-MP2.5 model,⁷⁴ linearized coupled-cluster doubles (DF-LCCD),⁷⁴ and recently for the density-fitted CCSD and CCD methods (DF-CCSD and DF-CCD).⁶⁴ However, one cannot compute analytic gradients with the CD integrals unless one uses an approach like that of Aquilante et al.⁷⁵

The computational advantages of the DF approach for analytic gradients are that one can replace the 4-index integrals and 4-index two-particle density matrix (TPDM) with 2-index and 3-index integrals and TPDMs, thus minimizing input/output (I/O) and storage requirements and reducing the computational cost of certain terms. Further, the DF approach also facilitates the solution of the Z-vector equation³¹ and the computation of the generalized Fock matrix (GFM).⁶⁰

Energies for the density-fitted and Cholesky decomposed CCSD(T) methods have been presented in previous studies.^{58,65,76–78} In this study, analytic gradient expressions are reported for the density-fitted CCSD(T) method [DF-CCSD(T)], where the density-fitting has been applied to the overall energy expression. The equations presented have been coded in c++ language and added to the Dfocc module^59–65 of the Psi4 package.⁷⁹ The computational time of the DF-CCSD(T) gradients is compared with that of the conventional CCSD(T) method (from Cfour).⁸⁰ The DF-CCSD(T) method is applied to equilibrium geometries and vibrational frequencies.

II. THEORETICAL APPROACH

A. Density-fitting

In the DF approximation, the atomic-orbital (AO) basis integrals are written as

{(μ ν | λ σ)}_{D F} = \sum_{Q}^{N_{a u x}} b_{μ ν}^{Q} b_{λ σ}^{Q},

(1)

and the DF factor $b_{μ ν}^{Q}$ is defined as

b_{μ ν}^{Q} = \sum_{P}^{N_{a u x}} (μ ν | P) {[J^{- 1 ∕ 2}]}_{P Q},

(2)

where

(μ ν | P) = \int \int χ_{μ} (r_{1}) χ_{ν} (r_{1}) \frac{1}{r_{12}} φ_{P} (r_{2}) d r_{1} d r_{2}

(3)

and

J_{P Q} = \int \int φ_{P} (r_{1}) \frac{1}{r_{12}} φ_{Q} (r_{2}) d r_{1} d r_{2},

(4)

where ${φ_{P} (r)}$ and ${χ_{μ} (r)}$ are auxiliary and primary basis functions, respectively.

Two-electron integrals (TEIs) over molecular-orbital (MO) can be written similarly as

{(p q | r s)}_{D F} = \sum_{Q}^{N_{a u x}} b_{p q}^{Q} b_{r s}^{Q},

(5)

where $b_{p q}^{Q}$ is a MO basis DF factor.

B. DF-CCSD energy and amplitude equations

At first, we employ the conventional notation for the orbital indices: i, j, k, l, m, n for occupied orbitals, a, b, c, d, e, f for virtual orbitals, and p, q, r, s, t, u for general spin orbitals. The DF-CCSD correlation energy can be expressed as follows:

Δ E = ⟨ 0 | e^{- \hat{T}} {\hat{H}}_{N} e^{\hat{T}} | 0 ⟩,

(6)

where ${\hat{H}}_{N}$ is the normal-ordered Hamiltonian operator,^4,81 $|) 0 (⟩$ is the reference determinant, and $\hat{T}$ is the cluster excitation operator. For the CCSD wave function,

\hat{T} = {\hat{T}}_{1} + {\hat{T}}_{2},

(7)

where ${\hat{T}}_{1}$ and ${\hat{T}}_{2}$ are single- and double-excitation operators, respectively,

{\hat{T}}_{1} = \sum_{i}^{o c c} \sum_{a}^{v i r} t_{i}^{a} {\hat{a}}^{†} î,

(8)

{\hat{T}}_{2} = \frac{1}{4} \sum_{i, j}^{o c c} \sum_{a, b}^{v i r} t_{i j}^{a b} {\hat{a}}^{†} {\hat{b}}^{†} ĵ î,

(9)

where ${\hat{a}}^{†}$ and $î$ are the creation and annihilation operators and $t_{i}^{a}$ and $t_{i j}^{a b}$ are the single and double excitation amplitudes, respectively.

The amplitude equations can be written as

⟨ Φ_{i}^{a} | e^{- \hat{T}} {\hat{H}}_{N} e^{\hat{T}} | 0 ⟩ = 0,

(10)

⟨ Φ_{i j}^{a b} | e^{- \hat{T}} {\hat{H}}_{N} e^{\hat{T}} | 0 ⟩ = 0,

(11)

where $⟨) Φ_{i}^{a} |$ and $⟨) Φ_{i j}^{a b} |$ are singly- and doubly-excited Slater determinants, respectively.

C. DF-CCSD- $Λ$ energy functional (unrelaxed Lagrangian)

It is convenient to introduce a Lagrangian^82,83 (DF-CCSD- $Λ$ functional) to obtain a variational energy functional (⁠ $\tilde{L}$ ⁠) as follows:

\tilde{L} = ⟨ 0 | (1 + \hat{Λ}) e^{- \hat{T}} Ĥ e^{\hat{T}} | 0 ⟩,

(12)

where $Ĥ$ is the Hamiltonian operator and $\hat{Λ}$ is the coupled-cluster de-excitation operator. For the CCSD wave function,

\hat{Λ} = {\hat{Λ}}_{1} + {\hat{Λ}}_{2},

(13)

where ${\hat{Λ}}_{1}$ and ${\hat{Λ}}_{2}$ are the single and double de-excitation operators, respectively,

{\hat{Λ}}_{1} = \sum_{i}^{o c c} \sum_{a}^{v i r} λ_{a}^{i} {\hat{i}}^{†} â,

(14)

{\hat{Λ}}_{2} = \frac{1}{4} \sum_{i, j}^{o c c} \sum_{a, b}^{v i r} λ_{a b}^{i j} {\hat{i}}^{†} {\hat{j}}^{†} \hat{b} â,

(15)

where $λ_{a}^{i}$ and $λ_{a b}^{i j}$ are the single and double de-excitation amplitudes, respectively.

The DF-CCSD T- and $Λ$ -amplitude equations^{35,37,38,84,85} are obtained using variational properties of $\tilde{L}$ ⁠,

⟨ 0 | (1 + \hat{Λ}) [e^{- \hat{T}} Ĥ e^{\hat{T}} - E] | Φ_{i}^{a} ⟩ = 0,

(16)

⟨ 0 | (1 + \hat{Λ}) [e^{- \hat{T}} Ĥ e^{\hat{T}} - E] | Φ_{i j}^{a b} ⟩ = 0,

(17)

where E is the DF-CCSD total energy.

In our previous study,⁶⁴ all explicit equations for DF-CCSD T and $Λ$ amplitudes and particle density matrices (PDMs) were provided in the spin-orbital formalism. Hence, we will not repeat those equations. However, spin-free equations (for T and $Λ$ amplitudes, and PDMs) for a closed-shell RHF reference based DF-CCSD method are reported in the supplementary material.

D. Triples energy correction for DF-CCSD

1. Spin-orbital equations

Following the prescription of Bartlett and co-workers,^3,5,6,86 the perturbative triples corrections for the DF-CCSD method⁶⁵ can be obtained from the following general equation:

E_{T}^{[4]} = ⟨ 0 | () {\hat{T}}_{1}^{†} + {\hat{T}}_{2}^{†})) {\hat{W}}_{N} {\hat{T}}_{3}^{[2]} | 0 ⟩ + ⟨ 0 | {\hat{T}}_{2}^{†} {\hat{f}}_{N}^{o} {\hat{T}}_{3}^{[2]} | 0 ⟩,

(18)

where ${\hat{f}}_{N}^{o}$ is the off-diagonal part of the normal-ordered Fock operator and ${\hat{W}}_{N}$ is the two-electron component of ${\hat{H}}_{N}$ ⁠. With the canonical HF orbitals, the second term of Eq. (18) vanishes. The second-order triples excitation operator is defined as follows:

{\hat{T}}_{3}^{[2]} = \frac{1}{36} \sum_{i j k}^{o c c} \sum_{a b c}^{v i r} t {(c)}_{i j k}^{a b c} {\hat{a}}^{†} {\hat{b}}^{†} {\hat{c}}^{†} \hat{k} ĵ î,

(19)

where $t {(c)}_{i j k}^{a b c}$ are the connected triple amplitudes.

It is convenient to express the (T) correction as follows:^22,23

E_{(T)} = \frac{1}{36} \sum_{i j k}^{o c c} \sum_{a b c}^{v i r} [t {(c)}_{i j k}^{a b c} + t {(d)}_{i j k}^{a b c}] D_{i j k}^{a b c} t {(c)}_{i j k}^{a b c},

(20)

where $t {(d)}_{i j k}^{a b c}$ are the disconnected triple amplitudes. $t {(c)}_{i j k}^{a b c}$ and $t {(d)}_{i j k}^{a b c}$ are defined as

\begin{array}{rcl} D_{i j k}^{a b c} t {(c)}_{i j k}^{a b c} & = & P (i j ∕ k) P (a b ∕ c) \sum_{e}^{v i r} t_{i j}^{a e} {⟨ b c | | e k ⟩}_{D F} \\ - P (i j ∕ k) P (a b ∕ c) \sum_{m}^{o c c} t_{i m}^{a b} {⟨ m c | | j k ⟩}_{D F}, \end{array}

(21)

\begin{array}{rcl} D_{i j k}^{a b c} t {(d)}_{i j k}^{a b c} & = & P (i j ∕ k) P (a b ∕ c) t_{k}^{c} {⟨ a b | | i j ⟩}_{D F} \\ + P (i j ∕ k) P (a b ∕ c) t_{i j}^{a b} f_{k c}, \end{array}

(22)

where

D_{i j k}^{a b c} = f_{i i} + f_{j j} + f_{k k} - f_{a a} - f_{b b} - f_{c c} .

(23)

The permutation operator is defined as

P (p q ∕ r) = 1 - P (p r) - P (q r),

(24)

where $P (p r)$ acts to permute indices p and r.

Then, we may define the combined triple amplitudes as follows:

t_{i j k}^{a b c} = t {(c)}_{i j k}^{a b c} + t {(d)}_{i j k}^{a b c} .

(25)

Finally, the (T) correction can be written as

E_{(T)} = \frac{1}{36} \sum_{i j k}^{o c c} \sum_{a b c}^{v i r} t_{i j k}^{a b c} D_{i j k}^{a b c} t {(c)}_{i j k}^{a b c} .

(26)

Further, we may also define the combined ${\hat{T}}_{3}$ operator as

{\hat{T}}_{3} = \frac{1}{36} \sum_{i j k}^{o c c} \sum_{a b c}^{v i r} t_{i j k}^{a b c} {\hat{a}}^{†} {\hat{b}}^{†} {\hat{c}}^{†} \hat{k} ĵ î .

(27)

2. Closed-shell spin-free equations

The closed-shell spin-free equation for the triples corrections can be written as follows:^87,88

E_{(T)} = \frac{1}{3} \sum_{i j k}^{o c c} \sum_{a b c}^{v i r} () 4 W_{i j k}^{a b c} + W_{i j k}^{b c a} + W_{i j k}^{c a b})) () V_{i j k}^{a b c} - V_{i j k}^{c b a})) D_{i j k}^{a b c},

(28)

where

\begin{array}{rcl} W_{i j k}^{a b c} = P_{i j k}^{a b c} (\sum_{e}^{v i r} t_{k j}^{c e} {(i a | b e)}_{D F} - \sum_{m}^{o c c} t_{i m}^{a b} {(m j | k c)}_{D F}) /) D_{i j k}^{a b c}, \end{array}

(29)

V_{i j k}^{a b c} = W_{i j k}^{a b c} + (t_{i}^{a} {(j b | k c)}_{D F} + t_{j}^{b} {(i a | k c)}_{D F} + t_{k}^{c} {(i a | j b)}_{D F}) /) D_{i j k}^{a b c},

(30)

and

P_{i j k}^{a b c} X_{i j k}^{a b c} = X_{i j k}^{a b c} + X_{i k j}^{a c b} + X_{j i k}^{b a c} + X_{j k i}^{b c a} + X_{k i j}^{c a b} + X_{k j i}^{c b a} .

(31)

E. Triples Lagrangian

For the triples correction, we introduce the following unrelaxed Lagrangian:

\begin{array}{rcl} L_{(T)} & = & ⟨ 0 | () {\hat{T}}_{1}^{†} + {\hat{T}}_{2}^{†})) {\hat{W}}_{N} {\hat{T}}_{3}^{[2]} | 0 ⟩ + ⟨ 0 | {\hat{T}}_{2}^{†} {\hat{f}}_{N}^{o} {\hat{T}}_{3}^{[2]} | 0 ⟩ \\ + ⟨ 0 | {\hat{Λ}}_{3} () {\hat{f}}_{N}^{d} {\hat{T}}_{3}^{[2]} + {\hat{W}}_{N} {\hat{T}}_{2})) | 0 ⟩, \end{array}

(32)

where

{\hat{Λ}}_{3} = \frac{1}{36} \sum_{i j k}^{o c c} \sum_{a b c}^{v i r} λ_{a b c}^{i j k} {\hat{i}}^{†} {\hat{j}}^{†} {\hat{k}}^{†} ĉ \hat{b} â .

(33)

Since the Lagrangian is stationary with respect to triples amplitudes, it can readily be shown that

λ_{a b c}^{i j k} = t_{i j k}^{a b c} .

(34)

Therefore, we may re-write the Lagrangian as follows:

\begin{array}{rcl} L_{(T)} & = & ⟨ 0 | () {\hat{T}}_{1}^{†} + {\hat{T}}_{2}^{†})) {\hat{W}}_{N} {\hat{T}}_{3}^{[2]} | 0 ⟩ + ⟨ 0 | {\hat{T}}_{2}^{†} {\hat{f}}_{N}^{o} {\hat{T}}_{3}^{[2]} | 0 ⟩ \\ + ⟨ 0 | {\hat{T}}_{3}^{†} () {\hat{f}}_{N}^{d} {\hat{T}}_{3}^{[2]} + {\hat{W}}_{N} {\hat{T}}_{2})) | 0 ⟩ . \end{array}

(35)

F. Triples contributions to $Λ$ -amplitudes

1. Spin-orbital equations

Differentiating the Lagrangian with respect to singles and doubles amplitudes, we obtain contributions of $L_{(T)}$ to $Λ$ -amplitudes, which are denoted as ${}^{(T)}λ_{a}^{i}$ and ${}^{(T)}λ_{a b}^{i j}$ ⁠. Hence, contributions of $L_{(T)}$ to $Λ_{1}$ -amplitudes can be written as

{}^{(T)}λ_{a}^{i} = \frac{1}{4} \sum_{j k}^{o c c} \sum_{b c}^{v i r} t {(c)}_{i j k}^{a b c} {⟨ j k | | b c ⟩}_{D F},

(36)

and their contributions to the overall amplitudes are

λ_{a}^{i} = {}^{D F - C C S D}λ_{a}^{i} + {}^{(T)}λ_{a}^{i} ∕ D_{i}^{a} .

(37)

Similarly, contributions of $L_{(T)}$ to $Λ_{2}$ -amplitudes can be written as

\begin{array}{rcl} {}^{(T)}λ_{a b}^{i j} & = & \frac{1}{2} P_{-} (a b) \sum_{e, f}^{v i r} \sum_{k}^{o c c} M_{i j k}^{a e f} {⟨ b k | | e f ⟩}_{D F} \\ - \frac{1}{2} P_{-} (i j) \sum_{m, n}^{o c c} \sum_{c}^{v i r} M_{i m n}^{a b c} {⟨ m n | | j c ⟩}_{D F} \\ + \sum_{k}^{o c c} \sum_{c}^{v i r} f_{k c} t {(c)}_{i j k}^{a b c}, \end{array}

(38)

where

M_{i j k}^{a b c} = 2 t {(c)}_{i j k}^{a b c} + t {(d)}_{i j k}^{a b c},

(39)

and their contributions to the overall amplitudes are

λ_{a b}^{i j} = {}^{D F - C C S D}λ_{a b}^{i j} + {}^{(T)}λ_{a b}^{i j} ∕ D_{i j}^{a b} .

(40)

In the paper of Watts, Gauss, and Bartlett,²² there is $t_{i j k}^{a b c}$ in Eq. (38) instead of $t {(c)}_{i j k}^{a b c}$ ⁠. However, this appears to be a typographical error.

2. Closed-shell spin-free equations

At first, we would like to note that the spin-free equations for $Λ$ -amplitudes and PDMs were reported by Scuseria⁴⁰ and by Lee and Rendell⁴¹ in terms of the so-called Z-amplitudes. However, their definition of Z₁- and Z₂-amplitudes and PDMs is different from ours. For example, the relation between Z₂-amplitudes and $Λ_{2}$ -amplitudes may be written as $Z_{i j}^{a b} = 4 λ_{a b}^{i j} - 2 λ_{a b}^{j i}$ ⁠. Therefore, instead of using their formulas, we spin-adapted our spin-orbital equations following the prescription of Scuseria and Schaefer⁸⁹ for spin-adaptation of triples amplitudes, which is not as trivial as double amplitudes. Hence, we obtain the following formula for triples contributions to $Λ_{1}$ -amplitudes:

\begin{array}{rcl} {}^{(T)}λ_{a}^{i} & = & \frac{1}{2} \sum_{j k}^{o c c} \sum_{b c}^{v i r} (j b | k c) () 4 W_{i j k}^{a b c} + W_{i j k}^{b c a} + W_{i j k}^{c a b} \\ - 3 W_{i j k}^{c b a} - 2 W_{i j k}^{a c b} - W_{i j k}^{b a c})) . \end{array}

(41)

Similarly, our spin-free equation for $Λ_{2}$ -amplitudes is

\begin{array}{rcl} {}^{(T)}λ_{a b}^{i j} & = & P_{+} (i a, j b) \sum_{e, f}^{v i r} \sum_{k}^{o c c} () 2 M_{i j k}^{a e f} - M_{i j k}^{f e a} - M_{i j k}^{a f e})) {(k f | b e)}_{D F} \\ - P_{+} (i a, j b) \sum_{m, n}^{o c c} \sum_{c}^{v i r} () 2 M_{i m n}^{a b c} - M_{i m n}^{c b a} - M_{i m n}^{a c b})) {(m j | n c)}_{D F} \\ + \sum_{k}^{o c c} \sum_{c}^{v i r} f_{k c} () 2 W_{i j k}^{a b c} - W_{i j k}^{c b a} - W_{i j k}^{a c b})), \end{array}

(42)

where

M_{i j k}^{a b c} = W_{i j k}^{a b c} + V_{i j k}^{a b c} .

(43)

G. Triples contributions to PDMs

1. Spin-orbital equations

Using our definitions of one- and two-particle density matrices (OPDM and TPDM),^60,64 we can write the following general equation for triples corrections for OPDM diagonal elements:

γ_{p}^{(c o r r)} = \frac{1}{2} P_{+} (p q) ⟨ 0 | {\hat{T}}_{3}^{†} {{\hat{p}}^{†} \hat{p}} {\hat{T}}_{3}^{[2]} | 0 ⟩ .

(44)

More explicitly, we have the following equations:

γ_{i}^{(c o r r)} = - \frac{1}{12} \sum_{j k}^{o c c} \sum_{a b c}^{v i r} t {(c)}_{i j k}^{a b c} t_{i j k}^{a b c},

(45)

γ_{a}^{(c o r r)} = \frac{1}{12} \sum_{i j k}^{o c c} \sum_{b c}^{v i r} t {(c)}_{i j k}^{a b c} t_{i j k}^{a b c} .

(46)

At this stage, we would like to note that in our previous studies,^60,64 we employed the non-canonical perturbed orbitals (NCPOs) to obtain our non-canonical gradient (NCG) equations. However, we may also use the canonical-perturbed orbitals (CPOs) to obtain a strictly canonical gradient (SCG) formalism. In the SCG, we need to compute only diagonal blocks of the unrelaxed OPDM, which is scaled as O(N⁵) for DF-CCSD. However, in the SCG, we need a more complex Z-vector equation compared with NCG. Hence, there is no advantage of SCG over NCG in the case of DF-CCSD. However, in the case of DF-CCSD(T), the SCG approach is very helpful since the scaling of full OPDM is O(N⁷). Thus, with the SCG approach, one needs only diagonal OPDMs, which are scaled as O(N⁶), instead of the full OPDM, which is scaled as O(N⁷).

The 3-index TPDM can be written as follows:

Γ_{i m}^{Q (c o r r)} = \frac{1}{2} P_{+} (i m) \sum_{j}^{o c c} \sum_{a}^{v i r} M_{i j a m} b_{j a}^{Q},

(47)

\begin{array}{rcl} Γ_{i a}^{Q (c o r r)} & = & \frac{1}{2} \sum_{j}^{o c c} \sum_{m}^{o c c} M_{j i a m} b_{j m}^{Q} + \frac{1}{2} \sum_{j}^{o c c} \sum_{b}^{v i r} M_{i j a b} b_{j b}^{Q} \\ + \frac{1}{2} \sum_{b}^{v i r} \sum_{d}^{v i r} M_{i a b d} b_{d b}^{Q}, \end{array}

(48)

Γ_{a b}^{Q (c o r r)} = \frac{1}{2} P_{+} (a b) \sum_{i}^{o c c} \sum_{c}^{v i r} M_{i c b a} b_{i c}^{Q},

(49)

where

M_{i j a m} = \frac{1}{2} \sum_{k}^{o c c} \sum_{b c}^{v i r} M_{i j k}^{a b c} t_{m k}^{b c},

(50)

M_{i j a b} = \sum_{k}^{o c c} \sum_{c}^{v i r} t {(c)}_{i j k}^{a b c} t_{k}^{c},

(51)

M_{i a b d} = \frac{1}{2} \sum_{j k}^{o c c} \sum_{c}^{v i r} M_{i j k}^{a b c} t_{j k}^{d c} .

(52)

Finally, we note that there is no difference between the NCG and SCG for TPDMs.

2. Closed-shell spin-free equations

Spin-free equations for the OPDM can be written as

\begin{array}{rcl} γ_{i}^{(c o r r)} & = & - \sum_{j k}^{o c c} \sum_{a b c}^{v i r} () 4 W_{i j k}^{a b c} + W_{j k i}^{a b c} + W_{k i j}^{a b c} - W_{k j i}^{a b c} \\ - 2 W_{i k j}^{a b c} - W_{j i k}^{a b c})) V_{i j k}^{a b c}, \end{array}

(53)

\begin{array}{rcl} γ_{a}^{(c o r r)} & = & \sum_{i j k}^{o c c} \sum_{b c}^{v i r} () 4 W_{i j k}^{a b c} + W_{i j k}^{c a b} + W_{i j k}^{b c a} - 3 W_{i j k}^{c b a} \\ - 2 W_{i j k}^{a c b} - W_{i j k}^{b a c})) V_{i j k}^{a b c} . \end{array}

(54)

Similarly, TPDMs can be written as

Γ_{i m}^{Q (c o r r)} = P_{+} (i m) \sum_{j}^{o c c} \sum_{a}^{v i r} M_{i j a m} b_{j a}^{Q},

(55)

\begin{array}{rcl} Γ_{i a}^{Q (c o r r)} & = & \sum_{j}^{o c c} \sum_{m}^{o c c} M_{j i a m} b_{j m}^{Q} + \sum_{j}^{o c c} \sum_{b}^{v i r} M_{i j a b} b_{j b}^{Q} \\ + \sum_{b}^{v i r} \sum_{d}^{v i r} M_{i a b d} b_{b d}^{Q}, \end{array}

(56)

Γ_{a b}^{Q (c o r r)} = P_{+} (a b) \sum_{i}^{o c c} \sum_{c}^{v i r} M_{i c a b} b_{i c}^{Q},

(57)

where

\begin{array}{rcl} M_{i j a m} & = & \sum_{k}^{o c c} \sum_{b c}^{v i r} () 2 M_{i j k}^{a b c} + 2 M_{i j k}^{c a b} + 2 M_{i j k}^{b c a} - M_{i j k}^{c b a} \\ - M_{i j k}^{a c b} - 4 M_{i j k}^{b a c})) t_{m k}^{b c}, \end{array}

(58)

\begin{array}{rcl} M_{i a b d} & = & \sum_{j k}^{o c c} \sum_{c}^{v i r} () 4 M_{i j k}^{a b c} + M_{i j k}^{c a b} + M_{i j k}^{b c a} - 2 M_{i j k}^{c b a} \\ - 2 M_{i j k}^{a c b} - 2 M_{i j k}^{b a c})) t_{j k}^{d c}, \end{array}

(59)

\begin{array}{rcl} M_{i j a b} & = & \sum_{k}^{o c c} \sum_{c}^{v i r} t_{k}^{c} () 4 W_{i j k}^{a b c} + W_{i j k}^{c a b} + W_{i j k}^{b c a} - 2 W_{i j k}^{c b a} \\ - 2 W_{i j k}^{a c b} - 2 W_{i j k}^{b a c})) . \end{array}

(60)

H. DF-CCSD(T) analytic gradients

For the DF-CCSD(T) wave function, the following Lagrangian can be employed (in the spin-orbital formalism):

L = ⟨ 0 | (1 + \hat{Λ}) e^{- \hat{T}} Ĥ e^{\hat{T}} | 0 ⟩ + \sum_{p q} z_{p q} f_{p q},

(61)

where the Z-vector³¹ satisfies

z_{p q} = z_{q p},

(62)

z_{p p} = 0 .

(63)

More explicitly,

\begin{array}{rcl} L & = & ⟨ 0 | (1 + \hat{Λ}) e^{- \hat{T}} Ĥ e^{\hat{T}} | 0 ⟩ + 2 \sum_{a i} z_{a i} f_{a i} + \sum_{i j} z_{i j} f_{i j} + \sum_{a b} z_{a b} f_{a b} . \end{array}

(64)

The solution of the Z-vector equations will be presented in Sec. II I.

The gradient of energy can be written as follows:^{82,83,90–96}

{\frac{d E}{d x} |)}_{x = x_{0}} = {\frac{\partial L}{\partial x} |)}_{x = x_{0}},

(65)

and when we insert the explicit expressions for integral derivatives into the above equation, the gradient equation can be cast into the following form:⁶⁰

\begin{array}{rcl} \frac{d E}{d x} & = & \sum_{p, q} γ_{p q}^{e f f} h_{p q}^{x} - \sum_{p, q} F_{p q}^{e f f} S_{p q}^{x} + \sum_{Q}^{N_{a u x}} \sum_{p q} {\tilde{Γ}}_{p q}^{Q (e f f)} {(Q | p q)}^{x} \\ - \sum_{P, Q}^{N_{a u x}} Γ_{P Q}^{e f f} J_{P Q}^{x}, \end{array}

(66)

where $γ_{p q}^{e f f}$ ⁠, $Γ_{P Q}^{e f f}$ ⁠, ${\tilde{Γ}}_{p q}^{Q (e f f)}$ ⁠, and $F_{p q}^{e f f}$ are the relaxed OPDM, 2- and 3-index TPDM, and GFM, respectively. The unrelaxed OPDM^25,97 is

γ_{p q} = \frac{1}{2} P_{+} (p q) ⟨ 0 | (1 + \hat{Λ}) e^{- \hat{T}} {\hat{p}}^{†} \hat{q} e^{\hat{T}} | 0 ⟩,

(67)

and the unrelaxed 2- and 3-index TPDMs are defined by^60,64

Γ_{p q}^{Q} = \frac{1}{2} {\hat{P}}_{+} (p q) \sum_{r, s} ⟨ 0 | (1 + \hat{Λ}) e^{- \hat{T}} {\hat{p}}^{†} {\hat{r}}^{†} ŝ \hat{q} e^{\hat{T}} | 0 ⟩ b_{r s}^{Q},

(68)

{\tilde{Γ}}_{p q}^{Q} = \sum_{P}^{N_{a u x}} Γ_{p q}^{P} {[J^{- 1 ∕ 2}]}_{P Q},

(69)

Γ_{P Q} = \frac{1}{2} \sum_{p, q} c_{p q}^{P} {\tilde{Γ}}_{p q}^{Q} = \frac{1}{2} \sum_{p, q} {\tilde{Γ}}_{p q}^{P} c_{p q}^{Q},

(70)

c_{p q}^{Q} = \sum_{P}^{N_{a u x}} b_{p q}^{P} {[J^{- 1 ∕ 2}]}_{P Q} .

(71)

Explicit equations for the unrelaxed GFM are provided in our previous studies.^59–61,64

In the SCG, we need to consider only diagonal elements of the unrelaxed OPDM. Hence, with the following definition:

γ_{p q}^{(c o r r)} = δ_{p q} γ_{p}^{(c o r r)},

(72)

we can use our previously reported formulas for the GFM and the separable part of the TPDM.⁶⁴

In the SCG, Z-vector contributions to the OPDM and the separable part of the TPDM can be written as follows:

γ_{p q}^{e f f} = γ_{p q} + z_{p q},

(73)

\begin{array}{rcl} Γ_{i j}^{Q (e f f)} & = & Γ_{i j}^{Q} + z_{i j} J_{Q} + δ_{i j} () Z_{Q} + Z_{Q}^{'} + Z_{Q}^{″})) - \frac{1}{2} () Z_{i j}^{Q} + Z_{j i}^{Q})) - Z_{i j}^{Q'}, \end{array}

(74)

Γ_{a i}^{Q (e f f)} = Γ_{i a}^{Q (e f f)} = Γ_{i a}^{Q} + z_{a i} J_{Q} - Z_{a i}^{Q} - Z_{a i}^{Q''},

(75)

Γ_{a b}^{Q (e f f)} = Γ_{a b}^{Q} + z_{a b} J_{Q} .

(76)

Similarly, the Z-vector contributions to the GFM can be written as follows:

F_{i a}^{e f f} = F_{i a} + 𝜀_{i} z_{a i},

(77)

F_{a b}^{e f f} = F_{a b} + 𝜀_{a} z_{a b},

(78)

\begin{array}{rcl} F_{i j}^{e f f} & = & F_{i j} + 𝜀_{i} z_{i j} + \sum_{Q}^{N_{a u x}} b_{i j}^{Q} Z_{Q} - \frac{1}{2} \sum_{Q}^{N_{a u x}} \sum_{m}^{o c c} b_{i m}^{Q} Z_{m j}^{Q} \\ - \sum_{Q}^{N_{a u x}} \sum_{e}^{v i r} b_{i e}^{Q} Z_{e j}^{Q} + \sum_{Q}^{N_{a u x}} b_{i j}^{Q} Z_{Q}^{'} - \frac{1}{2} \sum_{Q}^{N_{a u x}} \sum_{l}^{o c c} b_{i l}^{Q} Z_{l j}^{Q'} \\ + \sum_{Q}^{N_{a u x}} b_{i j}^{Q} Z_{Q}^{″} - \sum_{Q}^{N_{a u x}} \sum_{d}^{v i r} b_{i d}^{Q} Z_{d j}^{Q''}, \end{array}

(79)

\begin{array}{rcl} F_{a i}^{e f f} & = & F_{a i} + z_{a i} 𝜀_{a} + \sum_{Q}^{N_{a u x}} b_{a i}^{Q} Z_{Q} - \frac{1}{2} \sum_{Q}^{N_{a u x}} \sum_{m}^{o c c} b_{a m}^{Q} Z_{m i}^{Q} \\ - \sum_{Q}^{N_{a u x}} \sum_{e}^{v i r} b_{a e}^{Q} Z_{e i}^{Q} + \frac{1}{2} () w_{a i}^{'} + w_{a i}^{″})), \end{array}

(80)

where $𝜀_{p}$ are orbital energies, and intermediates are defined as

J_{Q} = \sum_{m}^{o c c} b_{m m}^{Q},

(81)

Z_{Q} = 2 \sum_{a i} z_{a i} b_{a i}^{Q},

(82)

Z_{a i}^{Q} = \sum_{m} z_{a m} b_{m i}^{Q},

(83)

Z_{i j}^{Q} = 2 \sum_{a} z_{a i} b_{a j}^{Q},

(84)

Z_{Q}^{'} = \sum_{k l} z_{k l} b_{k l}^{Q},

(85)

Z_{i j}^{Q'} = 2 \sum_{k} z_{i k} b_{k j}^{Q},

(86)

Z_{Q}^{″} = \sum_{c d} z_{c d} b_{c d}^{Q},

(87)

Z_{a i}^{Q''} = \sum_{b} z_{a b} b_{b i}^{Q},

(88)

w_{a i}^{'} = 2 \sum_{Q}^{N_{a u x}} b_{a i}^{Q} Z_{Q}^{'} - \sum_{Q}^{N_{a u x}} \sum_{l}^{o c c} b_{a l}^{Q} Z_{l i}^{Q'},

(89)

w_{a i}^{″} = 2 \sum_{Q}^{N_{a u x}} b_{a i}^{Q} Z_{Q}^{″} - 2 \sum_{Q}^{N_{a u x}} \sum_{d}^{v i r} b_{a d}^{Q} Z_{d i}^{Q''} .

(90)

We would like to note that for the Z-vector contributions to the PDMs and GFM, the $b_{p q}^{Q}$ integrals should employ the same auxiliary basis set (e.g., the JK-FIT basis set) as the density-fitted reference function, since these contributions originate from the Fock matrix derivatives.

The gradient is evaluated by back-transforming the relaxed PDMs and the relaxed GFM into the AO basis and contracting them against AO integral derivatives as usual,^98–100

F_{μ ν} = \sum_{p q} C_{μ p} C_{ν q} F_{p q}^{e f f},

(91)

γ_{μ ν} = \sum_{p q} C_{μ p} C_{ν q} γ_{p q}^{e f f},

(92)

Γ_{μ ν}^{Q} = \sum_{p q} C_{μ p} C_{ν q} {\tilde{Γ}}_{p q}^{Q (e f f)},

(93)

where $F_{μ ν}$ ⁠, $γ_{μ ν}$ ⁠, and $Γ_{μ ν}^{Q}$ are the AO basis GFM, OPDM, and 3-index TPDM, respectively. The 2-index TPDM can be written as

Γ_{P Q} = \frac{1}{2} \sum_{p, q} c_{p q}^{P} {\tilde{Γ}}_{p q}^{Q (e f f)} = \frac{1}{2} \sum_{μ ν} c_{μ ν}^{P} Γ_{μ ν}^{Q},

(94)

c_{μ ν}^{Q} = \sum_{P}^{N_{a u x}} b_{μ ν}^{P} {[J^{- 1 ∕ 2}]}_{P Q} .

(95)

Hence, the final gradient equation in the AO basis can be written as follows:⁶⁰

\begin{array}{rcl} \frac{d E}{d x} & = & \sum_{μ ν} γ_{μ ν} h_{μ ν}^{x} - \sum_{μ ν} F_{μ ν} S_{μ ν}^{x} + \sum_{Q}^{N_{a u x}} \sum_{μ ν} Γ_{μ ν}^{Q} {(Q | μ ν)}^{x} \\ - \sum_{P, Q}^{N_{a u x}} Γ_{P Q} J_{P Q}^{x} . \end{array}

(96)

I. SCG Z-vector equations

Before presenting Z-vector equations, let us define the orbital gradient as follows:

w_{p q} = 2 () F_{p q} - F_{q p})) .

(97)

We obtain a set of linear equations for Z-vectors by differentiating the Lagrangian in Eq. (61) with respect to orbital rotation parameters. We solve the Z-vector equations in three steps. At first, we obtain the OO-block Z-vector as follows:

z_{i j} = - \frac{w_{i j}}{2 () 𝜀_{i} - 𝜀_{j}))} .

(98)

Then, the VV-block Z-vector as

z_{a b} = - \frac{w_{a b}}{2 () 𝜀_{a} - 𝜀_{b}))} .

(99)

Finally, the VO-block Z-vector,

w_{a i}^{e f f} + \sum_{b j} A_{a i, b j} z_{b j} = 0,

(100)

where A is the HF MO Hessian, and

w_{a i}^{e f f} = w_{a i} + w_{a i}^{'} + w_{a i}^{″} .

(101)

A detailed discussion for the iterative solution of Eq. (100) is reported in our previous study.⁶⁰ Further, singularities in Eqs. (98) and (99), in the case of orbital degeneracies, are avoided applying the method suggested by Lee and Rendell.⁴¹

III. RESULTS AND DISCUSSION

Results from the CCSD(T) and DF-CCSD(T) methods were obtained for a set of alkanes for comparison of the computational cost for single-point analytic gradient computations, using geometries from our recent study.⁵⁹ For the alkanes, Dunning’s correlation-consistent polarized valence triple- $ζ$ (cc-pVTZ) basis set was employed with the frozen core approximation.^101,102 For the cc-pVTZ primary basis sets, cc-pVTZ-JKFIT⁴⁹ and cc-pVTZ-RI¹⁰³ auxiliary basis sets were employed for reference and correlation energies, respectively.

Additionally, the CCSD(T) and DF-CCSD(T) methods were applied to a set of small molecules¹⁰⁴ for comparison of geometries and vibrational frequencies. These computations used Dunning’s correlation-consistent polarized valence quadruple- $ζ$ basis set (cc-pVQZ) with the frozen core approximation.^101,102 For the cc-pVQZ primary basis sets, cc-pVQZ-JKFIT⁴⁹ and cc-pVQZ-RI¹⁰³ auxiliary basis sets were employed for reference and correlation energies, respectively. The DF-CCSD(T) computations were performed with our present code, which is available in the development version of the Psi4 program package,⁷⁹ while the conventional CCSD(T) computations were carried out with the Cfour1.0 program.⁸⁰ Energies were converged tightly to 10⁻⁸ hartree. We verified our DF-CCSD(T) analytic gradients code with respect to numerical gradients (with the 5-point formulas).

A. Efficiency of DF-CCSD(T) analytic gradients

We consider a set of alkanes⁵⁹ to assess the efficiency of the DF-CCSD(T) analytic gradients. The total computational (wall) times for single-point frozen-core analytic gradient computations using the CCSD(T) and DF-CCSD(T) methods are presented in Fig. 1. These computations were performed on a single core of an Intel(R) Xeon(R) central processing unit (CPU) E5-2620 v3 at 2.40 GHz computer [although we perform computations with just a single core for a clearer comparison between programs, our code has been partially parallelized, through basic linear algebra subprogram (BLAS) routines, to use multiple cores]. For each software, $\sim$ 64 GB memory is provided. Further, in CCSD(T) computations (with Cfour), the AO-basis algorithm is chosen for the particle-particle ladder (PPL) term with the ABCDTYPE = AOBASIS option. For the largest member of the alkane set, $C_{6} H_{14}$ ⁠, the number of active occupied orbitals, virtual orbitals, and auxiliary basis functions is O = 19, V = 351, and N_aux = 906, respectively.

FIG. 1.

View large Download slide

Total wall-time (in min) for computations of single-point frozen-core analytic gradients for the C_nH_2n+2 (n = 1–6) set from the DF-CCSD(T) and CCSD(T) (from Cfour) methods with the cc-pVTZ basis set. For the largest member, the number of basis functions is 376. All computations were performed with an 10⁻⁸ energy convergence tolerance on a single node (1 core) Intel(R) Xeon(R) CPU E5-2620 v3 at 2.40 GHz computer (memory $\sim$ 64 GB).

As shown in Fig. 1, the DF-CCSD(T) method significantly reduces the total computational cost compared to conventional CCSD(T); there is more than 2-fold reduction in wall time compared to CCSD(T) for the largest member of the alkane test set. Since we do not have detailed timings for different parts of the CCSD(T) gradient code in Cfour, we cannot compare results for different parts of the CCSD(T) gradient computations. However, our recent study⁶⁵ demonstrates that our DF-CCSD energy code is significantly faster (about 5-fold in the case of $C_{8} H_{18}$ ⁠) than the conventional CCSD code of Cfour. Further, another recent study⁶⁴ shows that for DF-CCSD [hence for DF-CCSD(T)], the computation of gradient related terms,¹⁰⁵ which was the most expensive part for CCSD gradients of the Molpro2012.1 program,^106,107 is substantially accelerated (more than 8-fold for $C_{10} H_{22}$ ⁠) with the help of DF approximations. Hence, due to dramatically reduced cost for gradient terms and significantly reduced I/O time, the DF approach enables substantially accelerated analytic gradient codes for CC methods.

Next, we analyze the cost of different parts of the DF-CCSD(T) analytic gradient computations. Figure 2 shows the total computational time broken down into four major components: (1) the total time to converge the DF-CCSD energy and amplitudes, (2) the total time for (T) correction and its contributions to the $Λ$ equations and PDMs, (3) the total time to converge the $Λ$ equations, and (4) the cost of the remaining gradient terms.¹⁰⁵ For the largest member of the alkane test set, $C_{6} H_{14}$ ⁠, 3.4% of the time is spent in the DF-CCSD energy and amplitude equations, 91.0% is spent in (T) correction and its contributions to the $Λ$ equations and PDMs, 4.9% is spent in the $Λ$ equations, and 0.7% is spent in the remaining gradient terms. For the remaining molecules, 23.6%–89.8% of the time is spent in the (T) part. Clearly the (T) correction part is the most cost expensive part of the DF-CCSD(T) gradient, as expected, for large molecules.

FIG. 2.

View large Download slide

Total wall-time (in min) and its major components, such as DF-CCSD, (T)-correction, $Λ$ -amplitudes, and gradient terms,¹⁰⁵ for computations of single-point frozen-core analytic gradients for the C_nH_2n+2 (n = 1–6) set from the DF-CCSD(T) method with the cc-pVTZ basis set. For the largest member, the number of basis functions is 376. All computations were performed with an 10⁻⁸ energy convergence tolerance on a single node (1 core) Intel(R) Xeon(R) CPU E5-2620 v3 at 2.40 GHz computer (memory $\sim$ 64 GB).

Finally, we illustrate the scaling with respect to the number of computer cores for the DF-CCSD(T) gradients for the alkane test set in Fig. 3. In our previous study,⁶⁴ it has been demonstrated that the DF-CCSD code scales fairly well for the large molecules. Hence, in this study we illustrate the scaling of the (T) part. However, we note that our present (T) gradient code is not fully parallelized yet. It is partially threaded through BLAS routines only. As shown in Fig. 3, our (T) gradient algorithm is partially parallelized.

FIG. 3.

View large Download slide

Ratio of total wall-time for single-point frozen-core analytic gradient computations with 2, 4, and 6 cores with respect to the serial mode for the $C_{4} H_{10}$ ⁠, $C_{5} H_{12}$ ⁠, and $C_{6} H_{14}$ molecules from the (T) part of the DF-CCSD(T) method with the cc-pVTZ basis set. All computations were performed with an 10⁻⁸ energy convergence tolerance on a single node (1 core) Intel(R) Xeon(R) CPU E5-2620 v3 at 2.40 GHz computer (memory $\sim$ 64 GB).

B. Accuracy of DF-CCSD(T) analytic gradients

To assess the accuracy of the DF-CCSD(T) method, we consider equilibrium geometries and harmonic vibrational frequencies. In order to assess the performance of DF-CCSD(T) for bond lengths, we consider a set of small molecules employed in previous studies.^104,108 Table I reports the bond lengths of the molecules considered in increasing order. Errors for DF-CCSD(T) with respect to conventional CCSD(T) are depicted in Fig. 4. The mean absolute error (⁠ $Δ_{m a e}$ ⁠), the standard deviation of errors (⁠ $Δ_{s t d}$ ⁠), and the maximum absolute error (⁠ $Δ_{m a x}$ ⁠) with respect to CCSD(T) are 0.0002, 0.0006, and 0.0028 Å, respectively. Further, we compare the performance of DF-CCSD(T) and CCSD(T) for bond lengths with respect to the experimental values.¹⁰⁸ $Δ_{m a e}$ values for DF-CCSD(T) and CCSD(T) are 0.0022 and 0.0023 Å, respectively, as depicted in Fig. 5. Hence, the performance of the CCSD(T) and DF-CCSD(T) methods is essentially the same, as in the case of CCSD/DF-CCSD⁶⁴ and MP2/DF-MP2.⁶⁰ Overall, the error introduced by the DF approximation is quite negligible.

TABLE I.

Bond lengths of molecules considered in order of increasing values (in Å).

	Molecule	Bond
1	HF	R_HF
2	$H_{2} O$	R_OH
3	$H_{2} O_{2}$	R_OH
4	HOF	R_OH
5	HNC	R_NH
6	${NH}_{3}$	R_NH
7	$N_{2} H_{2}$	R_NH
8	$C_{2} H_{2}$	R_CH
9	HNO	R_NH
10	HCN	R_CH
11	$C_{2} H_{4}$	R_CH
12	${CH}_{4}$	R_CH
13	$N_{2}$	R_NN
14	${CH}_{2} O$	R_CH
15	CO	R_CO
16	HCN	R_CN
17	${CO}_{2}$	R_CO
18	HNC	R_CN
19	$C_{2} H_{2}$	R_CC
20	${CH}_{2} O$	R_CO
21	HNO	R_NO
22	$N_{2} H_{2}$	R_NN
23	$C_{2} H_{4}$	R_CC
24	HOF	R_OF
25	$H_{2} O_{2}$	R_OO
26	BH	R_BH

	Molecule	Bond
1	HF	R_HF
2	$H_{2} O$	R_OH
3	$H_{2} O_{2}$	R_OH
4	HOF	R_OH
5	HNC	R_NH
6	${NH}_{3}$	R_NH
7	$N_{2} H_{2}$	R_NH
8	$C_{2} H_{2}$	R_CH
9	HNO	R_NH
10	HCN	R_CH
11	$C_{2} H_{4}$	R_CH
12	${CH}_{4}$	R_CH
13	$N_{2}$	R_NN
14	${CH}_{2} O$	R_CH
15	CO	R_CO
16	HCN	R_CN
17	${CO}_{2}$	R_CO
18	HNC	R_CN
19	$C_{2} H_{2}$	R_CC
20	${CH}_{2} O$	R_CO
21	HNO	R_NO
22	$N_{2} H_{2}$	R_NN
23	$C_{2} H_{4}$	R_CC
24	HOF	R_OF
25	$H_{2} O_{2}$	R_OO
26	BH	R_BH

FIG. 4.

View large Download slide

Errors in bond lengths of molecules considered (Table I) for the DF-CCSD(T) method with respect to CCSD(T) (cc-pVQZ basis set was employed).

FIG. 5.

View large Download slide

Mean absolute errors in bond lengths of molecules considered (Table I) for the CCSD(T) and DF-CCSD(T) methods with respect to experiment.

Next, we consider harmonic vibrational frequencies of the molecules shown in Table I. The $Δ_{m a e}$ ⁠, $Δ_{s t d}$ ⁠, and $Δ_{m a x}$ values are 0.7, 1.0, and 3.6 cm⁻¹, respectively. Thus, for harmonic vibrational frequencies, results of CCSD(T) and DF-CCSD(T) are essentially the same, similar to the CCSD/DF-CCSD case,⁶⁴ and the error introduced by the DF approach is again negligible.

IV. CONCLUSIONS

In this research, an efficient implementation of analytic gradients for the CCSD(T) method with the density-fitting (DF) approximation, which is denoted as DF-CCSD(T), has been reported. The computational time of the DF-CCSD(T) single point analytic gradients has been compared with that of the conventional CCSD(T) (from the Cfour1.0 package). The DF-CCSD(T) method significantly reduces the computational cost compared to CCSD(T), with a more than 2-fold reduction for the largest member of alkane test set considered, $C_{6} H_{14}$ in a cc-pVTZ basis set. Due to the dramatically reduced cost for the so-called gradient terms¹⁰⁵ and significantly reduced I/O time, the DF approach enables substantially accelerated analytic gradients for CC methods. Since four-dimensional electron repulsion integral and TPDM tensors can be avoided with the help of the DF approach, which leads to scaling reduction for certain terms and reduced I/O and storage requirements, the DF-CCSD(T) method is substantially more efficient than CCSD(T), especially for gradient terms.

DF-CCSD(T) has been applied for a variety of molecules to assess its accuracy for equilibrium bond lengths and harmonic vibrational frequencies. The DF approximation introduces negligible errors compared to CCSD(T): equilibrium bond lengths have a mean absolute error of 0.0002 Å (maximum error of 0.0028 Å), and harmonic vibrational frequencies exhibit mean absolute errors of 0.7 cm⁻¹ for the cases considered.

SUPPLEMENTARY MATERIAL

See supplementary material for spin-free equations (for T and $Λ$ amplitudes, and PDMs) for a closed-shell RHF reference based DF-CCSD method.

ACKNOWLEDGMENTS

This research was supported by the Scientific and Technological Research Council of Turkey (No. TÜB $\dot{I}$ TAK-114Z786), the European Cooperation in Science and Technology (No. CM1405), and the U.S. National Science Foundation (Grant Nos. ACI-1147843 and CHE-1566192). U.B. also acknowledges the support from Turkish Academy of Sciences, Outstanding Young Scientist Award (No. TÜBA-GEB $\dot{I}$ P 2015).

REFERENCES

1.

J.

Čížek

,

J. Chem. Phys.

45

,

4256

(

1966

).

https://doi.org/10.1063/1.1727484

Google Scholar

Crossref

2.

R. J.

Bartlett

,

Annu. Rev. Phys. Chem.

32

,

359

(

1981

).

https://doi.org/10.1146/annurev.pc.32.100181.002043

Google Scholar

Crossref

3.

R. J.

Bartlett

,

J. Phys. Chem.

93

,

1697

(

1989

).

https://doi.org/10.1021/j100342a008

Google Scholar

Crossref

4.

T. D.

Crawford

and

H. F.

Schaefer

,

Rev. Comput. Chem.

14

,

33

(

2000

).

https://doi.org/10.1002/9780470125915.ch2

Google Scholar

Crossref

5.

R. J.

Bartlett

and

M.

Musiał

,

Rev. Mod. Phys.

79

,

291

(

2007

).

https://doi.org/10.1103/revmodphys.79.291

Google Scholar

Crossref

6.

R. J.

Bartlett

,

Wiley Interdiscip. Rev.: Comput. Mol. Sci.

2

,

126

(

2012

).

https://doi.org/10.1002/wcms.76

Google Scholar

Crossref

7.

G. D.

Purvis

and

R. J.

Bartlett

,

J. Chem. Phys.

76

,

1910

(

1982

).

https://doi.org/10.1063/1.443164

Google Scholar

Crossref

8.

R. J.

Barlett

,

H.

Sekino

, and

G. D.

Purvis

,

Chem. Phys. Lett.

98

,

66

(

1983

).

https://doi.org/10.1016/0009-2614(83)80204-8

Google Scholar

Crossref

9.

Y. S.

Lee

,

S. A.

Kucharski

,

R. J.

Bartlett

, and

G. D.

Purvis

,

J. Chem. Phys.

81

,

5906

(

1984

).

https://doi.org/10.1063/1.447591

Google Scholar

Crossref

10.

J. A.

Pople

,

M.

Head-Gordon

, and

K.

Raghavachari

,

J. Chem. Phys.

87

,

5968

(

1987

).

https://doi.org/10.1063/1.453520

Google Scholar

Crossref

11.

K.

Raghavachari

,

G. W.

Trucks

,

J. A.

Pople

, and

M.

Head-Gordon

,

Chem. Phys. Lett.

157

,

479

(

1989

).

https://doi.org/10.1016/s0009-2614(89)87395-6

Google Scholar

Crossref

12.

R. J.

Bartlett

,

J. D.

Watts

,

S. A.

Kucharski

, and

J.

Noga

,

Chem. Phys. Lett.

165

,

513

(

1990

).

https://doi.org/10.1016/0009-2614(90)87031-l

Google Scholar

Crossref

13.

G. E.

Scuseria

and

T. J.

Lee

,

J. Chem. Phys.

93

,

5851

(

1990

).

https://doi.org/10.1063/1.459684

Google Scholar

Crossref

14.

G. E.

Scuseria

,

T. P.

Hamilton

, and

H. F.

Schaefer

,

J. Chem. Phys.

92

,

568

(

1990

).

https://doi.org/10.1063/1.458407

Google Scholar

Crossref

15.

M.

Urban

,

J.

Noga

,

S. J.

Cole

, and

R. J.

Bartlett

,

J. Chem. Phys.

83

,

4041

(

1985

).

https://doi.org/10.1063/1.449067

Google Scholar

Crossref

16.

T. J.

Lee

and

G. E.

Scuseria

, in

Quantum Mechanical Electronic Structure Calculations with Chemical Accuracy

, edited by

S. R.

Langhoff

(

Kluwer Academic

,

Dordrecht

,

1995

), pp.

47

–

108

.

Google Scholar

Crossref

17.

J. D.

Watts

,

J. F.

Stanton

, and

R. J.

Bartlett

,

Chem. Phys. Lett.

178

,

471

(

1991

).

https://doi.org/10.1016/0009-2614(91)87004-u

Google Scholar

Crossref

18.

G. E.

Scuseria

,

Chem. Phys. Lett.

176

,

423

(

1991

).

https://doi.org/10.1016/0009-2614(91)90231-w

Google Scholar

Crossref

19.

J.

Noga

, and

R. J.

Bartlett

,

J. Chem. Phys.

86

,

7041

(

1987

).

https://doi.org/10.1063/1.452353

Google Scholar

Crossref

20.

J. D.

Watts

and

R. J.

Bartlett

,

J. Chem. Phys.

96

,

6073

(

1992

).

https://doi.org/10.1063/1.462649

Google Scholar

Crossref

21.

J. R.

Thomas

,

B. J.

DeLeeuw

,

G.

Vacek

,

T. D.

Crawford

,

Y.

Yamaguchi

, and

H. F.

Schaefer

,

J. Chem. Phys.

99

,

403

(

1993

).

https://doi.org/10.1063/1.465764

Google Scholar

Crossref

22.

J. D.

Watts

,

J.

Gauss

, and

R. J.

Bartlett

,

J. Chem. Phys.

98

,

8718

(

1993

).

https://doi.org/10.1063/1.464480

Google Scholar

Crossref

23.

T. D.

Crawford

and

H. F.

Schaefer

,

J. Chem. Phys.

104

,

6259

(

1996

).

https://doi.org/10.1063/1.471287

Google Scholar

Crossref

24.

T. D.

Crawford

,

T. J.

Lee

, and

H. F.

Schaefer

,

J. Chem. Phys.

107

,

7943

(

1997

).

https://doi.org/10.1063/1.475081

Google Scholar

Crossref

25.

U.

Bozkaya

and

H. F.

Schaefer

,

J. Chem. Phys.

136

,

204114

(

2012

).

https://doi.org/10.1063/1.4720382

Google Scholar

Crossref

PubMed

26.

P.

Pulay

,

Mol. Phys.

17

,

197

(

1969

).

https://doi.org/10.1080/00268976900100941

Google Scholar

Crossref

27.

P.

Pulay

, in

Applications of Electronic Structure Theory

, edited by

H. F.

Schaefer

(

Springer

,

Boston

,

1977

), pp.

153

–

185

.

Google Scholar

Crossref

28.

P.

Pulay

,

Theor. Chem. Acc.

50

,

299

(

1979

).

https://doi.org/10.1007/bf00551337

Google Scholar

Crossref

29.

P.

Pulay

,

G.

Fogarasi

,

F.

Pang

, and

J. E.

Boggs

,

J. Am. Chem. Soc.

101

,

2550

(

1979

).

https://doi.org/10.1021/ja00504a009

Google Scholar

Crossref

30.

J.

Gauss

, in

Modern Methods and Algorithms of Quantum Chemistry: Proceedings

, NIC Series, edited by

J.

Grotendorst

(

John von Neumann Institute for Computing

,

Jülich

,

2000

), pp.

541

–

592

.

Google Scholar

31.

N. C.

Handy

and

H. F.

Schaefer

,

J. Chem. Phys.

81

,

5031

(

1984

).

https://doi.org/10.1063/1.447489

Google Scholar

Crossref

32.

L.

Adamowicz

,

W. D.

Laidig

, and

R. J.

Bartlett

,

Int. J. Quantum Chem.

26

(

S18

),

245

(

1984

).

https://doi.org/10.1002/qua.560260825

Google Scholar

Crossref

33.

G.

Fitzgerald

,

R.

Harrison

,

W. D.

Laidig

, and

R. J.

Bartlett

,

Chem. Phys. Lett.

117

,

433

(

1985

).

https://doi.org/10.1016/0009-2614(85)80278-5

Google Scholar

Crossref

34.

A. C.

Scheiner

,

G. E.

Scuseria

,

J. E.

Rice

,

T. J.

Lee

, and

H. F.

Schaefer

,

J. Chem. Phys.

87

,

5361

(

1987

).

https://doi.org/10.1063/1.453655

Google Scholar

Crossref

35.

E. A.

Salter

,

G. W.

Trucks

, and

R. J.

Bartlett

,

J. Chem. Phys.

90

,

1752

(

1989

).

https://doi.org/10.1063/1.456069

Google Scholar

Crossref

36.

A. P.

Rendell

and

T. J.

Lee

,

J. Chem. Phys.

94

,

6219

(

1991

).

https://doi.org/10.1063/1.460410

Google Scholar

Crossref

37.

J.

Gauss

,

J. F.

Stanton

, and

R. J.

Bartlett

,

J. Chem. Phys.

95

,

2623

(

1991

).

https://doi.org/10.1063/1.460915

Google Scholar

Crossref

38.

J.

Gauss

,

J. F.

Stanton

, and

R. J.

Bartlett

,

J. Chem. Phys.

95

,

2639

(

1991

).

https://doi.org/10.1063/1.460916

Google Scholar

Crossref

39.

J.

Gauss

,

W. J.

Lauderdale

,

J. F.

Stanton

,

J. D.

Watts

, and

R. J.

Bartlett

,

Chem. Phys. Lett.

182

,

207

(

1991

).

https://doi.org/10.1016/0009-2614(91)80203-a

Google Scholar

Crossref

40.

G. E.

Scuseria

,

J. Chem. Phys.

94

,

442

(

1991

).

https://doi.org/10.1063/1.460359

Google Scholar

Crossref

41.

T. J.

Lee

and

A. P.

Rendell

,

J. Chem. Phys.

94

,

6229

(

1991

).

https://doi.org/10.1063/1.460411

Google Scholar

Crossref

42.

J. D.

Watts

,

J.

Gauss

, and

R. J.

Bartlett

,

Chem. Phys. Lett.

200

,

1

(

1992

).

https://doi.org/10.1016/0009-2614(92)87036-o

Google Scholar

Crossref

43.

K.

Hald

,

A.

Halkier

,

P.

Jørgensen

,

S.

Coriani

,

C.

Hättig

, and

T.

Helgaker

,

J. Chem. Phys.

118

,

2985

(

2003

).

https://doi.org/10.1063/1.1531106

Google Scholar

Crossref

44.

J. L.

Whitten

,

J. Chem. Phys.

58

,

4496

(

1973

).

https://doi.org/10.1063/1.1679012

Google Scholar

Crossref

45.

B. I.

Dunlap

,

J. W. D.

Connolly

, and

J. R.

Sabin

,

J. Chem. Phys.

71

,

3396

(

1979

).

https://doi.org/10.1063/1.438728

Google Scholar

Crossref

46.

M.

Feyereisen

,

G.

Fitzgerald

, and

A.

Komornicki

,

Chem. Phys. Lett.

208

,

359

(

1993

).

https://doi.org/10.1016/0009-2614(93)87156-w

Google Scholar

Crossref

47.

O.

Vahtras

,

J.

Almlöf

, and

M. W.

Feyereisen

,

Chem. Phys. Lett.

213

,

514

(

1993

).

https://doi.org/10.1016/0009-2614(93)89151-7

Google Scholar

Crossref

48.

A. P.

Rendell

and

T. J.

Lee

,

J. Chem. Phys.

101

,

400

(

1994

).

https://doi.org/10.1063/1.468148

Google Scholar

Crossref

49.

F.

Weigend

,

Phys. Chem. Chem. Phys.

4

,

4285

(

2002

).

https://doi.org/10.1039/b204199p

Google Scholar

Crossref

50.

A.

Sodt

,

J. E.

Subotnik

, and

M.

Head-Gordon

,

J. Chem. Phys.

125

,

194109

(

2006

).

https://doi.org/10.1063/1.2370949

Google Scholar

Crossref

PubMed

51.

H.-J.

Werner

,

F. R.

Manby

, and

P. J.

Knowles

,

J. Chem. Phys.

118

,

8149

(

2003

).

https://doi.org/10.1063/1.1564816

Google Scholar

Crossref

52.

M.

Schütz

and

F. R.

Manby

,

Phys. Chem. Chem. Phys.

5

,

3349

(

2003

).

https://doi.org/10.1039/b304550a

Google Scholar

Crossref

53.

H.-J.

Werner

and

M.

Schütz

,

J. Chem. Phys.

135

,

144116

(

2011

).

https://doi.org/10.1063/1.3641642

Google Scholar

Crossref

PubMed

54.

N. H. F.

Beebe

and

J.

Linderberg

,

Int. J. Quantum Chem.

12

,

683

(

1977

).

https://doi.org/10.1002/qua.560120408

Google Scholar

Crossref

55.

I.

Røeggen

and

E.

Wisløff-Nilssen

,

Chem. Phys. Lett.

132

,

154

(

1986

).

https://doi.org/10.1016/0009-2614(86)80099-9

Google Scholar

Crossref

56.

H.

Koch

,

A. S.

de Meras

, and

T. B.

Pedersen

,

J. Chem. Phys.

118

,

9481

(

2003

).

https://doi.org/10.1063/1.1578621

Google Scholar

Crossref

57.

F.

Aquilante

,

T. B.

Pedersen

, and

R.

Lindh

,

J. Chem. Phys.

126

,

194106

(

2007

).

https://doi.org/10.1063/1.2736701

Google Scholar

Crossref

PubMed

58.

A. E.

DePrince

and

C. D.

Sherrill

,

J. Chem. Theory Comput.

9

,

2687

(

2013

).

https://doi.org/10.1021/ct400250u

Google Scholar

Crossref

PubMed

59.

U.

Bozkaya

,

J. Chem. Theory Comput.

10

,

2371

(

2014

).

https://doi.org/10.1021/ct500231c

Google Scholar

Crossref

PubMed

60.

U.

Bozkaya

,

J. Chem. Phys.

141

,

124108

(

2014

).

https://doi.org/10.1063/1.4896235

Google Scholar

Crossref

PubMed

61.

U.

Bozkaya

,

J. Chem. Theory Comput.

10

,

4389

(

2014

).

https://doi.org/10.1021/ct500634s

Google Scholar

Crossref

PubMed

62.

U.

Bozkaya

,

J. Chem. Theory Comput.

12

,

1179

(

2016

).

https://doi.org/10.1021/acs.jctc.5b01128

Google Scholar

Crossref

PubMed

63.

U.

Bozkaya

,

Phys. Chem. Chem. Phys.

18

,

11362

(

2016

).

https://doi.org/10.1039/c6cp00164e

Google Scholar

Crossref

PubMed

64.

U.

Bozkaya

and

C. D.

Sherrill

,

J. Chem. Phys.

144

,

174103

(

2016

).

https://doi.org/10.1063/1.4948318

Google Scholar

Crossref

PubMed

65.

U.

Bozkaya

,

J. Chem. Phys.

144

,

144108

(

2016

).

https://doi.org/10.1063/1.4945706

Google Scholar

Crossref

PubMed

66.

F.

Weigend

and

M.

Häser

,

Theor. Chem. Acc.

97

,

331

(

1997

).

https://doi.org/10.1007/s002140050269

Google Scholar

Crossref

67.

C.

Hättig

,

A.

Hellweg

, and

A.

Köhn

,

Phys. Chem. Chem. Phys.

8

,

1159

(

2006

).

https://doi.org/10.1039/b515355g

Google Scholar

Crossref

PubMed

68.

R. A.

Distasio

,

R. P.

Steele

,

Y. M.

Rhee

,

Y.

Shao

, and

M.

Head-Gordon

,

J. Comput. Chem.

28

,

839

(

2007

).

https://doi.org/10.1002/jcc.20604

Google Scholar

Crossref

PubMed

69.

C.

Hättig

,

J. Chem. Phys.

118

,

7751

(

2003

).

https://doi.org/10.1063/1.1564061

Google Scholar

Crossref

70.

A.

Köhn

and

C.

Hättig

,

J. Chem. Phys.

119

,

5021

(

2003

).

https://doi.org/10.1063/1.1597635

Google Scholar

Crossref

71.

M.

Schütz

,

H.-J.

Werner

,

R.

Lindh

, and

F. R.

Manby

,

J. Chem. Phys.

121

,

737

(

2004

).

https://doi.org/10.1063/1.1760747

Google Scholar

Crossref

PubMed

72.

W.

Györffy

,

T.

Shiozaki

,

G.

Knizia

, and

H.-J.

Werner

,

J. Chem. Phys.

138

,

104104

(

2013

).

https://doi.org/10.1063/1.4793737

Google Scholar

Crossref

PubMed

73.

K.

Ledermüller

and

M.

Schütz

,

J. Chem. Phys.

140

,

164113

(

2014

).

https://doi.org/10.1063/1.4872169

Google Scholar

Crossref

PubMed

74.

U.

Bozkaya

, available in the Psi4 package, “

Density-fitted orbital-optimized coupled-cluster (DFOCC) module

” (unpublished) .

75.

F.

Aquilante

,

R.

Lindh

, and

T. B.

Pedersen

,

J. Chem. Phys.

129

,

034106

(

2008

).

https://doi.org/10.1063/1.2955755

Google Scholar

Crossref

PubMed

76.

E.

Epifanovsky

,

D.

Zuev

,

X.

Feng

,

K.

Khistyaev

,

Y.

Shao

, and

A. I.

Krylov

,

J. Chem. Phys.

139

,

134105

(

2013

).

https://doi.org/10.1063/1.4820484

Google Scholar

Crossref

PubMed

77.

J.

Boström

,

M.

Pitoňák

,

F.

Aquilante

,

P.

Neogrády

,

T. B.

Pedersen

, and

R.

Lindh

,

J. Chem. Theory Comput.

8

,

1921

(

2012

).

https://doi.org/10.1021/ct3003018

Google Scholar

Crossref

PubMed

78.

A. E.

DePrince

,

M. R.

Kennedy

,

B. G.

Sumpter

, and

C. D.

Sherrill

,

Mol. Phys.

112

,

844

(

2014

).

https://doi.org/10.1080/00268976.2013.874599

Google Scholar

Crossref

79.

R. M.

Parrish

,

L. A.

Burns

,

D. G. A.

Smith

,

A. C.

Simmonett

,

A. E.

DePrince

,

E. G.

Hohenstein

,

U.

Bozkaya

,

A. Y.

Sokolov

,

R. D.

Remigio

,

R. M.

Richard

,

J. F.

Gonthier

,

A. M.

James

,

H. R.

McAlexander

,

A.

Kumar

,

M.

Saitow

,

X.

Wang

,

B. P.

Pritchard

,

P.

Verma

,

H. F.

Schaefer

,

K.

Patkowski

,

R. A.

King

,

E. F.

Valeev

,

F. A.

Evangelista

,

J. M.

Turney

,

T. D.

Crawford

, and

C. D.

Sherrill

,

J. Chem. Theory Comput.

13

,

3185

(

2017

).

https://doi.org/10.1021/acs.jctc.7b00174

Google Scholar

Crossref

PubMed

80.

CFOUR, a quantum chemical program package written by

J. F.

Stanton

,

J.

Gauss

,

M. E.

Harding

, and

P. G.

Szalay

with contributions from

A. A.

Auer

,

R. J.

Bartlett

,

U.

Benedikt

,

C.

Berger

,

D. E.

Bernholdt

,

Y. J.

Bomble

,

L.

Cheng

,

O.

Christiansen

,

M.

Heckert

,

O.

Heun

,

C.

Huber

,

T.-C.

Jagau

,

D.

Jonsson

,

J.

Jusélius

,

K.

Klein

,

W. J.

Lauderdale

,

D. A.

Matthews

,

T.

Metzroth

,

D. P.

O’Neill

,

D. R.

Price

,

E.

Prochnow

,

K.

Ruud

,

F.

Schiffmann

,

W.

Schwalbach

,

S.

Stopkowicz

,

A.

Tajti

,

J.

Vázquez

,

F.

Wang

, and

J. D.

Watts

and the integral packages MOLECULE (

J.

Almlöf

and

P. R.

Taylor

), PROPS (

P. R.

Taylor

), ABACUS (

T.

Helgaker

,

H. J. Aa.

Jensen

,

P.

Jørgensen

, and

J.

Olsen

), and ECP routines by

A. V.

Mitin

and

C.

van Wüllen

.

81.

I.

Shavitt

and

R. J.

Bartlett

,

Many-Body Methods in Chemistry and Physics

, 1st ed. (

Cambridge Press

,

New York

,

2009

), pp.

443

–

449

.

Google Scholar

Crossref

82.

T.

Helgaker

and

P.

Jørgensen

,

Adv. Quantum Chem.

19

,

183

(

1988

).

https://doi.org/10.1016/s0065-3276(08)60616-4

Google Scholar

Crossref

83.

P.

Jørgensen

and

T.

Helgaker

,

J. Chem. Phys.

89

,

1560

(

1988

).

https://doi.org/10.1063/1.455152

Google Scholar

Crossref

84.

J.

Gauss

and

J. F.

Stanton

,

J. Chem. Phys.

103

,

3561

(

1995

).

https://doi.org/10.1063/1.470240

Google Scholar

Crossref

85.

J.

Gauss

and

J. F.

Stanton

,

J. Chem. Phys.

116

,

1773

(

2001

).

https://doi.org/10.1063/1.1429244

Google Scholar

Crossref

86.

R. J.

Bartlett

,

Mol. Phys.

108

,

2905

(

2010

).

https://doi.org/10.1080/00268976.2010.531773

Google Scholar

Crossref

87.

T. J.

Lee

,

A. P.

Rendell

, and

P. R.

Taylor

,

J. Phys. Chem.

94

,

5463

(

1990

).

https://doi.org/10.1021/j100377a008

Google Scholar

Crossref

88.

A. P.

Rendell

,

T. J.

Lee

, and

A.

Komornicki

,

Chem. Phys. Lett.

178

,

462

(

1991

).

https://doi.org/10.1016/0009-2614(91)87003-t

Google Scholar

Crossref

89.

G. E.

Scuseria

and

H. F.

Schaefer

,

Chem. Phys. Lett.

152

,

382

(

1988

).

https://doi.org/10.1016/0009-2614(88)80110-6

Google Scholar

Crossref

90.

T.

Helgaker

,

P.

Jørgensen

, and

N.

Handy

,

Theor. Chem. Acc.

76

,

227

(

1989

).

https://doi.org/10.1007/bf00532006

Google Scholar

Crossref

91.

T.

Helgaker

and

P.

Jørgensen

,

Theor. Chem. Acc.

75

,

111

(

1989

).

https://doi.org/10.1007/bf00527713

Google Scholar

Crossref

92.

T. U.

Helgaker

and

J.

Almlöf

,

Int. J. Quantum Chem.

26

,

275

(

1984

).

https://doi.org/10.1002/qua.560260211

Google Scholar

Crossref

93.

T. U.

Helgaker

, in

Geometrical Derivatives of Energy Surfaces and Molecular Properties

, edited by

P.

Jørgensen

and

J.

Simons

(

Springer, Reidel

,

Dordrecht

,

1986

), pp.

1

–

16

.

Google Scholar

94.

J.

Simons

,

T. U.

Helgaker

, and

P.

Jørgensen

,

Chem. Phys.

86

,

413

(

1984

).

https://doi.org/10.1016/0301-0104(84)80029-4

Google Scholar

Crossref

95.

R.

Shepard

, in

Modern Electronic Structure Theory Part I

, 1st ed., Advanced Series in Physical Chemistry, edited by

D. R.

Yarkony

(

World Scientific Publishing Company

,

London

,

1995

), Vol. 2, pp.

345

–

458

.

Google Scholar

Crossref

96.

T.

Helgaker

, in

The Encyclopedia of Computational Chemistry

, edited by

P. R.

Schleyer

,

N. L.

Allinger

,

T.

Clark

,

J.

Gasteiger

,

P. A.

Kollman

,

H. F.

Schaefer

, and

P. R.

Schreiner

(

Wiley

,

Chichester

,

1998

), pp.

1157

–

1169

.

Google Scholar

97.

U.

Bozkaya

,

J. M.

Turney

,

Y.

Yamaguchi

,

H. F.

Schaefer

, and

C. D.

Sherrill

,

J. Chem. Phys.

135

,

104103

(

2011

).

https://doi.org/10.1063/1.3631129

Google Scholar

Crossref

PubMed

98.

J. E.

Rice

and

R. D.

Amos

,

Chem. Phys. Lett.

122

,

585

(

1985

).

https://doi.org/10.1016/0009-2614(85)87275-4

Google Scholar

Crossref

99.

Y.

Yamaguchi

,

Y.

Osamura

,

J. D.

Goddard

, and

H. F.

Schaefer

,

A New Dimension to Quantum Chemistry: Analytic Derivative Methods in Ab Initio Molecular Electronic Structure Theory

(

Oxford University Press

,

New York

,

1994

), pp.

29

–

52

and

128

–

143

.

Google Scholar

100.

Y.

Yamaguchi

and

H. F.

Schaefer

, in

Handbook of High-Resolution Spectroscopies

, edited by

M.

Quack

and

F.

Merkt

(

John Wiley & Sons

,

2011

), pp.

325

–

362

.

Google Scholar

101.

T. H.

Dunning

,

J. Chem. Phys.

90

,

1007

(

1989

).

https://doi.org/10.1063/1.456153

Google Scholar

Crossref

102.

D. E.

Woon

and

T. H.

Dunning

,

J. Chem. Phys.

103

,

4572

(

1995

).

https://doi.org/10.1063/1.470645

Google Scholar

Crossref

103.

F.

Weigend

,

A.

Köhn

, and

C.

Hättig

,

J. Chem. Phys.

116

,

3175

(

2002

).

https://doi.org/10.1063/1.1445115

Google Scholar

Crossref

104.

U.

Bozkaya

and

C. D.

Sherrill

,

J. Chem. Phys.

139

,

054104

(

2013

).

https://doi.org/10.1063/1.4816628

Google Scholar

Crossref

PubMed

105.

By gradient terms we mean the computations of PDMs, GFM, solution of the Z-vector equation, formation of the relaxed PDMs and GFM, back-transformation of relaxed PDMs and GFM to the AO basis, and evaluation of gradients in the AO basis. All these tasks are performed by our general DF gradient code,⁶⁰ which is called Dfgrad.

106.

H.-J.

Werner

,

P. J.

Knowles

,

G.

Knizia

,

F. R.

Manby

,

M.

Schütz

,

P.

Celani

,

T.

Korona

,

R.

Lindh

,

A.

Mitrushenkov

,

G.

Rauhut

,

K. R.

Shamasundar

,

T. B.

Adler

,

R. D.

Amos

,

A.

Bernhardsson

,

A.

Berning

,

D. L.

Cooper

,

M. J. O.

Deegan

,

A. J.

Dobbyn

,

F.

Eckert

,

E.

Goll

,

C.

Hampel

,

A.

Hesselmann

,

G.

Hetzer

,

T.

Hrenar

,

G.

Jansen

,

C.

Köppl

,

Y.

Liu

,

A. W.

Lloyd

,

R. A.

Mata

,

A. J.

May

,

S. J.

McNicholas

,

W.

Meyer

,

M. E.

Mura

,

A.

Nicklass

,

D. P.

O’Neill

,

P.

Palmieri

,

D.

Peng

,

K.

Pflüger

,

R.

Pitzer

,

M.

Reiher

,

T.

Shiozaki

,

H.

Stoll

,

A. J.

Stone

,

R.

Tarroni

,

T.

Thorsteinsson

, and

M.

Wang

, molpro, version 2012.1, a package of ab initio programs,

2012

, see http://www.molpro.net.

107.

H.-J.

Werner

,

P. J.

Knowles

,

G.

Knizia

,

F. R.

Manby

, and

M.

Schütz

,

Wiley Interdiscip. Rev.: Comput. Mol. Sci.

2

,

242

(

2012

).

https://doi.org/10.1002/wcms.82

Google Scholar

Crossref

108.

T.

Helgaker

,

J.

Gauss

,

P.

Jørgensen

, and

J.

Olsen

,

J. Chem. Phys.

106

,

6430

(

1997

).

https://doi.org/10.1063/1.473634

Google Scholar

Crossref

2017

Author(s)

Analytic energy gradients for the coupled-cluster singles and doubles with perturbative triples method with the density-fitting approximation

I. INTRODUCTION

II. THEORETICAL APPROACH

A. Density-fitting

B. DF-CCSD energy and amplitude equations

C. DF-CCSD- $Λ$ energy functional (unrelaxed Lagrangian)

D. Triples energy correction for DF-CCSD

1. Spin-orbital equations

2. Closed-shell spin-free equations

E. Triples Lagrangian

F. Triples contributions to $Λ$ -amplitudes

1. Spin-orbital equations

2. Closed-shell spin-free equations

G. Triples contributions to PDMs

1. Spin-orbital equations

2. Closed-shell spin-free equations

H. DF-CCSD(T) analytic gradients

I. SCG Z-vector equations

III. RESULTS AND DISCUSSION

A. Efficiency of DF-CCSD(T) analytic gradients

B. Accuracy of DF-CCSD(T) analytic gradients

IV. CONCLUSIONS

SUPPLEMENTARY MATERIAL

ACKNOWLEDGMENTS

REFERENCES

Supplementary Material

Citing articles via

Submit your article

Sign up for alerts

Resources

Explore

pubs.aip.org

Connect with AIP Publishing

Analytic energy gradients for the coupled-cluster singles and doubles with perturbative triples method with the density-fitting approximation

I. INTRODUCTION

II. THEORETICAL APPROACH

A. Density-fitting

B. DF-CCSD energy and amplitude equations

C. DF-CCSD-Λ energy functional (unrelaxed Lagrangian)

D. Triples energy correction for DF-CCSD

1. Spin-orbital equations

2. Closed-shell spin-free equations

E. Triples Lagrangian

F. Triples contributions to Λ-amplitudes

1. Spin-orbital equations

2. Closed-shell spin-free equations

G. Triples contributions to PDMs

1. Spin-orbital equations

2. Closed-shell spin-free equations

H. DF-CCSD(T) analytic gradients

I. SCG Z-vector equations

III. RESULTS AND DISCUSSION

A. Efficiency of DF-CCSD(T) analytic gradients

B. Accuracy of DF-CCSD(T) analytic gradients

IV. CONCLUSIONS

SUPPLEMENTARY MATERIAL

ACKNOWLEDGMENTS

REFERENCES

Supplementary Material

Citing articles via

Submit your article

Sign up for alerts

Resources

Explore

pubs.aip.org

Connect with AIP Publishing

This Feature Is Available To Subscribers Only

C. DF-CCSD- $Λ$ energy functional (unrelaxed Lagrangian)

F. Triples contributions to $Λ$ -amplitudes