Fractional-occupation-number based divide-and-conquer coupled-cluster theory

Fractional-occupation-number based divide-and-conquer coupled-cluster theory

Accepted Manuscript Research paper Fractional-occupation-number based divide-and-conquer coupled-cluster theory Takeshi Yoshikawa, Hiromi Nakai PII: D...

1MB Sizes 0 Downloads 41 Views

Accepted Manuscript Research paper Fractional-occupation-number based divide-and-conquer coupled-cluster theory Takeshi Yoshikawa, Hiromi Nakai PII: DOI: Reference:

S0009-2614(18)30787-5 https://doi.org/10.1016/j.cplett.2018.09.056 CPLETT 35968

To appear in:

Chemical Physics Letters

Received Date: Accepted Date:

4 August 2018 24 September 2018

Please cite this article as: T. Yoshikawa, H. Nakai, Fractional-occupation-number based divide-and-conquer coupled-cluster theory, Chemical Physics Letters (2018), doi: https://doi.org/10.1016/j.cplett.2018.09.056

This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.

Fractional-occupation-number based divide-and-conquer coupled-cluster theory

Takeshi Yoshikawaa, Hiromi Nakaia-c *

a

Department of Chemistry and Biochemistry, School of Advanced Science and

Engineering, Waseda University, 3-4-1 Okubo, Shinjuku-ku, Tokyo 169-8555, Japan b

Research Institute for Science and Engineering, Waseda University, 3-4-1 Okubo,

Shinjuku-ku, Tokyo 169-8555, Japan c

Elements Strategy Initiative for Catalysts and Batteries (ESICB), Kyoto University,

Katsura, Kyoto 615-8520, Japan _____________________________________________________________________ *

Corresponding author. FAX: +81-3-3205-2504, E-mail address: [email protected] (H. Nakai). URL: http://www.chem.waseda.ac.jp/nakai/ (H. Nakai).

1

Abstract We have extended the divide-and-conquer (DC) coupled-cluster with singles and doubles (CCSD) to a fractional occupation number (FON) formalism, denoted as FON-DC-CCSD, using the thermal Wick theorem. The motivation is to address the inconsistency in the treatment of orbital occupations between the DC-based Hartree– Fock and the DC-CCSD methods, which adopt the Fermi distribution function and the step function for orbital occupation, respectively. Numerical applications involving polyene chains and single-walled carbon nanotubes confirm that the proposed FON-DC-CCSD method reduces both energy errors and computational costs compared with the conventional DC-CCSD method.

2

1. Introduction The coupled-cluster (CC) theory is a robust theoretical approach that systematically improves the accuracy of calculation of the total energy and the wavefunction of atoms and molecules by considering the increase in the order of electron excitation. Based on the correlated method, the energetics of various molecules can be predicted with chemical accuracy. The CC method with single, double, and perturbative triple excitations (CCSD(T)) has been widely used in practical applications and benchmark studies over the years [1,2]. However, its high scaling with respect to the system size N remains a severe impediment for large systems. For example, the computational costs of CCSD and CCSD(T) are asymptotically O(N6) and O(N7), respectively. Many researchers have reported a wide variety of linear-scaling methods based on fragmentation approaches [3-19], which partition the system under consideration into several subsystems. Li et al. [9] categorized the fragmentation approaches into energy-based approaches [3-14] and density matrix (DM)-based approaches [15-29]. Energy-based approaches directly evaluate the total energy of the system from subsystem contributions. The approach is generally applicable to various theory levels such as Hartree–Fock (HF), Møller–Plesset perturbation, and CC theories. In the

3

DM-based approaches, the total energy is calculated from the assembled DM, which is constructed from the DMs of a series of subsystems. One of the features of the DM-based approaches is that a non-integer number of electrons can be treated for each subsystem. Thus, it is possible to reasonably treat charge- and/or spin-delocalized systems using the DM-based approach. Gordon et al. [30] recognized one-step and two-step techniques for the energy-based and DM-based approaches, respectively. The divide-and-conquer (DC) method proposed by Yang and Lee [15,16] can be directly applied to the density functional theory, and is an example of a DM-based approach. These authors investigated the performance of this approach for calculations including HF exchange interactions [20-22]. Furthermore, DC-based correlation methods [23-29] were developed by exploiting energy density analysis (EDA) techniques [31]. The correlation energy of the total system is evaluated by summing up subsystem contributions. Therefore, the DC-based correlation methods are categorized as an energy-based approach. Note that the subsystem orbitals are obtained by diagonalizing the subsystem Fock matrices, which are constructed using the assembled DM. However, the occupation numbers of the subsystem orbitals need to be approximated from non-integer to integer, to solve the subsystem correlation problem. Recently, Kobayashi and Taketsugu proposed the DC-based second-order Møller–

4

Plesset perturbation (DC-MP2) method based on the thermal Wick theorem [32-34] and improved its accuracy [35]. In this study, we have extended the fractional occupation number (FON) technique to the DC-CCSD theory, denoted as FON-DC-CCSD. The rest of this report is organized as follows: in Section 2, we present the theoretical aspects; Section 3 describes the numerical applications; Section 4 presents the conclusion.

5

2. Theory 2.1. DC-HF method This subsection provides a brief explanation of the DC-HF method, the results of which are the basis for the DC-CCSD calculations described in the following subsection. For simplicity, we consider the case of closed-shell systems. The extension to open-shell systems is straightforward. In the DC-HF method, the system under consideration is spatially divided into disjoint subsystems, which are called the central regions. A set of atomic orbitals (AOs) corresponding to a central region α is denoted by S (  . To improve the description of the subsystem, the area adjacent to the central region, which is called the buffer region, is considered during the expansion of the subsystem’s molecular orbitals (MOs). A set of AOs corresponding to the buffer region of the subsystem α, which is denoted by B(  , is added to S (  to construct a set of AOs in the localization region of the

subsystem α, L(  , as follows: L(   S(  B  .

(1)

The DC-HF method evaluates the MOs C  and the orbital energies   of the subsystem α by solving the following local Roothaan–Hall equation for the localization region:

6

F  C  S C   ,

(2)

where Sα and Fα represent the local overlap and Fock matrices, respectively, which are extracted from the total matrices. The DM of the entire system, D, is constructed from the local DM for the subsystem α, Dα, as follows:  . D   P D

(3)



Pα represents the partition matrix with elements given by

1  P  1 2  0

   S     S       B     S        S     B   

(4)

otherwise.

The local DM for the subsystem α, Dα, is obtained using the subsystem MOs Cα, the subsystem orbital energies ɛα, and the common Fermi level  F  D  2 f    F   p  Cp*Cp ,

(5)

f  x   1  exp  x  .

(6)

p

1

Here, indices {μ, ν} and {p, q} refer to AOs and MOs, respectively. Unlike the standard HF method, the DM is constructed on the basis of the Fermi distribution function fβ(x) for the inverse temperature    kBT 

1

instead of the step function θ(x). Note that the

Fermi distribution function is identical to the step function in the limit of β → ∞, i.e., T → 0 K. As the occupation numbers of the subsystem orbitals are non-integer as

7

shown in Fig. 1(a), it is possible to reasonably treat charge- and/or spin-delocalized systems. The common Fermi level  F is determined by a constraint on the total number of electrons in the entire system, N e , as follows:

Ne   D S .

(7)

 ,

Similarly, the number of electrons for the subsystem α, N eL  , is determined by

NeL  =

  D

 , L 

 

S ,

(8)

whose value is a non-integer because of fractional occupation. Note that N e is not consistent with the sum N eL  because of double counting of the number of electrons in the buffer regions. The partition matrix Pα in Eq. (4) avoids such double counting. From Eqs. (3) and (7),  Ne   P D S   N eS   . 

 ,



(9)

Here, N eS   is defined and rewritten as follows:

NeS   =

S  

Thus, N e

  P

 , L 

 

D S 

    D

S   L 

 

S .

(10)

is regarded as the number of electrons in the central region. Notably, the S  

summation of N e

is consistent with the total number of electrons in the entire S  

system, i.e., integer, although N e

is a non-integer.

8

2.2. DC-CCSD method This subsection briefly summarizes the conventional DC-CCSD method. As the DC-HF method yields the MOs C  and the orbital energies   of the subsystem α defined by L(  , the local CCSD calculations can be performed for the localization region. The conventional DC-CCSD method approximately separates the occupied orbitals { i , j  , …} and virtual orbitals { a , b , …} by replacing the Fermi distribution function f    F   p  with the step function   F   p  . The DC-based correlation energy is estimated by summing the correlation energies corresponding to the individual subsystems. To avoid double counting of the correlation energy owing to the overlap of the buffer regions, EDA is adopted to extract the correlation energy for the central region. DC Ecorr   Ecorr  .

S 

(11)



S   Ecorr 

where

C     S  

  i , j a ,b

* i





 j a b 2ti j ,a b  ti j ,b a ,

(12)

pq rs represents a two-electron integral notation. For the DC-CCSD method,

the two-electron excitation coefficient for the subsystem, t     , becomes i j ,a b

ti j ,a b  ti j ,a b  ti ,a t j ,b ,

(13)

where t   and t     are the single and double excitation amplitudes of the local i j ,a b i ,a

9

CCSD, respectively. The local DM in this treatment is given as follows:  D  2  F   p  Cp*Cp  2 Ci*Ci . occ

p

(14)

i

Thus, the number of electrons in the localization region L(  is evaluated as

NeL  =

L  

Note that N e

  D

 , L 

 

S .

(15)

is an integer. Furthermore, the number of electrons corresponding to

the central region S (  is given in a similar way to Eq. (10)

NeS   =

  P

 , L 

 

D S 

S  

Notably, the summation of N e

    D

S   L 

 

S .

(16)

over the subsystems is not consistent with the total

number of electrons in the entire system

Ne   NeS   . 

(17)

2.3. FON-DC-CCSD method This subsection presents a novel DC-CCSD method by adopting the FON formalism, i.e., FON-DC-CCSD. This formulation aims to address the inconsistency of the orbital occupations of the subsystem between the DC-HF and the conventional DC-CCSD methods—namely, the former treats the FON by utilizing the Fermi distribution function, whereas the latter approximately uses the integer occupation

10

number with the step function. Table 1 shows a comparison between the conventional and FON-type CCSD methods. In the FON formalism, arbitrary orbitals {p, q, r, s, …} possess the FONs of electrons and holes given by

f p  f   F   p  ,

(18)

f p  1  f   F   p  .

(19)

Thus, the occupied and virtual orbital spaces are no longer separable. The excitation operators are defined as

ˆp ,q  aq†a p ,

(20)

ˆpq ,rs  ar†as†aq a p ,

(21)

where a †p and a p represent the creation and annihilation operators, respectively. The non-zero commutation and anti-commutation relations become

 a†p , aq   f p pq ,

(22)

 a p , aq†   f p pq . 

(23)

Consequently, the thermal Wick theorem is given by the following non-zero contractions:

a†p aq  f p pq ,

11

(24)

a p aq†  f p pq .

(25)

With this theorem, the energy and amplitude equations for the FON-CCSD are derived as shown in Refs. 36 and 37. As in the case of the conventional DC-CCSD method, the correlation energy of the FON-DC-CCSD approach is estimated by summing the correlation energies for the central region using EDA. Instead of Eq. (12), the correlation energy in the case of the FON-DC-CCSD is evaluated as S   Ecorr 

   S  

 p

   

Cp*  q r  s f p f q f r f s

q r s

 2t

f p  f   F   p  ,

 p q ,r s

(26)





 t p q ,s r ,

f p  1  f   F   p  ,

(27)

where indices { p , q , r  , s } refer to the arbitrary orbitals of the subsystem α. As the numbers of electrons in L(  and S (  are evaluated using Eqs. (8) and (10), respectively, the FON-DC-CCSD method satisfies the constraint on the total number of electrons and the DC-HF method. A comparison between Eqs. (12) and (26) reveals that the computational cost



L   2 L increases significantly from O Nocc  N vir  2







L   4   to O N orb , where N occ , N vir , and

 represent the numbers of occupied, virtual, and arbitrary orbitals of the subsystem N orb

α, respectively. If the FONs of electrons and holes are close to zero, the contribution to

12

the correlation energy in Eq. (24) is negligible. Therefore, by introducing a threshold λ, we define the quasi-occupied and quasi-virtual orbitals that satisfy f p   and f p   , respectively. The present study adopts   1015 . The numbers of

quasi-occupied and quasi-virtual orbitals are represented as

  and N qvir , N qocc

      respectively. Although the relations Nocc and N vir are  Nqocc  Norb  Nqvir  Norb

    satisfied, Nocc and N vir are established in practical cases, as shown in  Nqocc  Nqvir

the following section. Consequently, the correlation energy for the central region in the FON-DC-CCSD case becomes 

Ecorr 

  Nqocc Nqvir

C     S  

 i









* i





 j a b fi f j f a fb 2ti j ,a b  ti j ,b a .

, j a ,b

(28)

The aforementioned FON-DC-CCSD method was implemented in the house-code of the GAMESS program [38].

13

3. Illustrative applications 3.1. Polyene chain In this section, the results of a numerical examination of the performance of the FON-DC-CCSD method in comparison with the conventional DC-CCSD calculations are presented. The examined systems were bond-alternating (BA) and the uniform (U) polyene chains, C40H42, are as shown in Fig. 2. All the  C-C-H and  C-C-C bond angles were fixed at 120°, and the C-H bond length was set at 1.07 Å. For the BA chain, the single-bond C-C and double-bond C=C lengths were set as 1.46 and 1.36 Å, respectively, whereas for the U chain, a constant C-C length of 1.40 Å was adopted. The utilized basis set was 6-31G** for C and H [40,41]. For all the CCSD calculations, MOs corresponding to the C 1s orbitals were treated as frozen cores. For the DC calculations, the temperature parameter of the Fermi function was fixed to 1578 K, which was a default value in the GAMESS program [38]. Individual C2H2(3) units were used as central regions, which were labeled from the left edge as shown in Fig. 2. The six units on the left and right sides were treated as the HF buffer region for the DC-HF calculations. By applying the dual-buffer scheme [39], the electron correlation was treated with different buffer sizes, namely, from one to six units on the left and right sides. In the dual-buffer scheme, we redetermined the Fermi level to conserve the total

14

number of electrons. Although the DM obtained using DC-CCSD is different from that obtained using DC-HF, the number of electrons within the localization region is conserved through the DC-correlation calculations. Table 2 shows the numbers of electrons in the localization and central regions obtained using the conventional and FON-type DC-CCSD calculations with a buffer size nb = 2, namely, { N eL  , N eS   } and { N eL  , N eS   }, respectively. Table 2 numerically confirms that N eL  becomes an integer as theoretically predicted, although N eL  is a non-integer. However, in the case of the BA chain, the differences between N eL  and N eL  are small, i.e., less than 0.218. The U chain leads to large differences between N eL  and N eL  , i.e., a maximum of 1.965. This is due to the delocalization nature of the electronic structure of the U chain. The formal numbers of electrons in the central regions, C2H2 for #2-#19 and C2H3 for #1 and #20, are 14 and 15, respectively. Although both N eS   and N eS   are non-integers, the corresponding values for the BA chain are considerably closer to the formal number of electrons, i.e., the differences are 0.027 for N eS   and 0.014 for N eS   at the maximum. In the case of the U chain, the discrepancies are in the range

-0.054 to 0.282 for N eS   and -0.155 to 0.095 for N eS   . The summation of N eS   is consistent with the total number of electrons in the entire system for both the BA and U

15

chains, which satisfies the theoretical condition. However, the summation of N eS   is not consistent with these results. In particular, in the case of U chain, the difference is 3.054. Figure 3 shows the buffer-size dependence of the conventional and FON-type DC-CCSD energy errors of C40H42 from the standard CCSD energy. In all cases, the errors decrease for larger buffer sizes. For the BA polyene chain, the errors of the conventional and FON-type DC-CCSD exhibit a fast convergence to zero: for nb ≥ 2, 1 kcal/mol or less, i.e., the so-called chemical accuracy. The difference in the energy errors between the conventional and FON-type DC-CCSD is small except for nb = 1. For the U chain, the errors of conventional DC-CCSD calculations slowly converge to zero, whereas the errors of the FON-DC-CCSD results show a fast convergence. Table 3 shows the buffer-size dependence of the numbers of (quasi-)occupied and (quasi-)virtual orbitals at subsystem #10, which is located in the middle of the polyene    chain. Although N occ and N vir in DC-CCSD are well defined, N qocc and N qvir in

  FON-DC-CCSD change owing to the FONs. The increase from N occ to N qocc as well

 as from N vir to N qvir results in an increase in the computational cost. However, in the

polyene chains, the increases are extremely small: namely, {+1, +1} and {+3, +3} for    { Nqocc , Nqvir  Nocc  N vir } of the BA and U chains, respectively. In Table 3, the number

16

 of CCSD iterations, Niter , which is also related to the computational cost, is shown for

 each case. For the BA chain, Niter in FON-DC-CCSD is close to that in DC-CCSD.

 However, Niter in FON-DC-CCSD is considerably smaller than that in DC-CCSD. This

indicates that the FON treatment improves the poor convergence owing to the inconsistency of the orbital occupations of the subsystem between the DC-HF and DC-CCSD calculations, in particular, for delocalized systems. Finally, we examine the computational cost of the conventional and FON-type DC-CCSD calculations. Figure 4 shows the energy errors from the standard CCSD results versus the central processing unit (CPU) time for the conventional and FON-type DC-CCSD calculations of the BA and U polyene chains, using the CPU times over eight cores of two Intel Xeon X5680 (3.33 GHz) processors. Each plot corresponds to the case of nb = 1, 2, 3, 4, and 5. At the same buffer size, it is determined that the CPU times of the FON-DC-CCSD calculations are typically smaller than those of the conventional  DC-CCSD. This is because the effect of the decrease in Niter is greater than that of the

  increases in N qocc and N qvir in terms of the computational cost. Furthermore, at the

same accuracy, i.e., the same energy error, the effectiveness of FON-DC-CCSD is apparent, especially, for the delocalized U chain.

17

3.2. Single-walled carbon nanotubes This subsection examines the applicability of the FON-DC-CCSD method to single-walled carbon nanotubes (SWCNTs). SWCNTs have different characteristics depending on the chiral vectors Ch (n,m). SWCNTs are metallic if mod(n – m, 3) = 0, and semiconducting if mod(n – m, 3) = 1 or 2. The examined systems were metallic and semiconducting SWCNTs with chiral vectors (3,3) and (3,2), whose lengths corresponded to 13 benzene units, terminated by hydrogen atoms. It is difficult to accurately treat SWCNTs using standard fragmentation methods, because the aromaticity results in complicated subsystems. Generally, fragmentation with chemical intuition might achieve higher accuracy. However, in the present study, we adopted the automatic fragmentation technique, which defines the subsystem, i.e., the central region, as a cubic box with a certain length without any chemical intuition. We used 3.0 Å for the cubic box as illustrated in Fig. S1 in the supporting information. Based on the dual-buffer scheme, only the electron correlation was treated with the DC approach after the conventional HF calculations. The atoms within rb of each subsystem are treated as the correlation buffer region. The numbers of electrons in the localization and central regions calculated using the conventional and FON-type DC-CCSD methods are summarized in Table S1 in the supporting information. A similar tendency to that of the

18

polyene chains was observed, as shown in Table 2. Figure 5 shows the buffer-radius dependence rb [Å] of the conventional and FON-type DC-CCSD energy errors of the metallic and semiconducting SWCNTs from the standard CCSD energy. Although the subsystems are heterogeneous, the exhibited errors converge to zero for larger buffer sizes. For the semiconducting SWCNT, the difference in the energy errors between the conventional and FON-type DC-CCSD calculations is small. For the metallic SWCNTs, the FON-DC-CCSD calculations show a rapid decrease in the energy error compared with the conventional DC-CCSD calculations.

19

Conclusion In this study, we applied the FON formalism to address the inconsistency in the treatment of orbital occupations between the DC-HF and DC-CCSD methods, where the FONs of the subsystem orbitals are kept from the DC-HF with FON-DC-CCSD. Numerical applications to polyene chains and SWCNTs confirmed that the proposed FON-DC-CCSD method reduces the energy errors from the standard CCSD results in comparison with the conventional DC-CCSD method. Consequently, smaller buffer sizes can be adopted in the case of the FON-DC-CCSD method. Although the occupied and virtual spaces slightly increase, a difference in the maximum of three orbitals is observed through an examination of the polyene chains. The computational times for the proposed FON-DC-CCSD calculations are less than those for the conventional DC-CCSD calculations, because the numbers of CCSD iteration are considerably reduced. The aforementioned characteristics are more apparent for systems with delocalized electronic structures, for which difficulties are generally encountered when applying fragmentation methods. Consequently, the proposed FON-DC-CCSD method can expand the applicability of fragmentation methods.

20

Acknowledgments Some of the presented calculations were performed at the Research Center for Computational Science (RCCS), Okazaki Research Facilities, Institutes of Natural Sciences (NINS). This study was supported by the program “Elements Strategy Initiative to Form Core Research Center” of the Ministry of Education, Culture, Sports, Science and Technology (MEXT).

21

References [1]

K. Raghavachari, G. W. Trucks, J. A. Pople, M. Head-Gordon, Chem. Phys. Lett. 157 (1989) 479.

[2]

T. J. Lee, G. E. Scuseria, Quantum Mechanical Electronic Structure Calculations with Chemical Accuracy, Kluwer Academic Publishers: Dordrecht (1995).

[3]

H. Stoll, Chem. Phys. Lett. 191 (1992) 548.

[4]

K. Kitaura, E. Ikeo, T. Asada, T. Nakano, M. Uebayashi, Chem. Phys. Lett. 313 (1999) 701.

[5]

S. Li, J. Ma, Y. Jiang, J. Comput. Chem. 23 (2002) 237.

[6]

D. W. Zhang, J. Z. H. Zhang, J. Chem. Phys. 119 (2003) 3599.

[7]

K. Babu, S. R. Gadre, J. Comput. Chem. 24 (2003) 484.

[8]

V. Deev, M. A. Collins, J. Chem. Phys. 112 (2005) 154102.

[9]

W. Li, S. Li, Y. S. Jiang, J. Phys. Chem. A 111 (2007) 2193.

[10] X. He, B. Wang, K. M. Merz, J. Phys. Chem. B 113 (2009) 10380. [11] N. J. Mayhall, K. Raghavachari, J. Chem. Theor. Comput. 7 (2011) 1336. [12] H. A. Le, H. J. Tan, J. F. Ouyang, R. P. A. Bettens, J. Chem. Theor. Comput. 8 (2012) 469. [13] R. M. Richard, J. M. Herbert, J. Chem. Phys. 137 (2012) 064113.

22

[14] J. J. Eriksen, P. Baudin, P. Ettenhuber, K. Kristensen, T. Kjærgaard, P. Jørgensen, J. Chem. Theor. Comput. 11 (2015) 2984. [15] W. Yang, Phys. Rev. Lett. 66 (1991) 1438. [16] W. Yang, T. S. Lee, J. Chem. Phys. 103 (1995) 5674. [17] T. E. Exner, P. G. Mezey, J. Phys. Chem. A 108 (2004) 4301. [18] X. He, J. Z. H. Zhang, J. Chem. Phys. 122 (2005) 031103. [19] W. Li, S. Li, J. Chem. Phys. 122 (2005) 194109. [20] T. Akama, M. Kobayashi, H. Nakai, J. Comput. Chem. 28 (2007) 2003. [21] T. Akama, M. Kobayashi, H. Nakai, Int. J. Quantum Chem. 109 (2009) 2706. [22] T. Akama, A. Fujii, M. Kobayashi, H. Nakai, Mol. Phys. 105 (2010) 2799. [23] M. Kobayashi, Y. Imamura, H. Nakai, J. Chem. Phys. 127 (2007) 074103. [24] M. Kobayashi, H. Nakai, J. Chem. Phys. 129 (2008) 044103. [25] M. Kobayashi, H. Nakai, J. Chem. Phys. 131 (2009) 114108. [26] T. Yoshikawa, M. Kobayashi, H. Nakai, Theor. Chem. Acc. 130 (2011) 411. [27] T. Yoshikawa, M. Kobayashi, H. Nakai, Int. J. Quant. Chem. 113 (2013) 218. [28] J. Seino, H. Nakai, J. Chem. Phys. 139 (2013) 034109. [29] M. Nakano, T. Yoshikawa, S. Hirata, J. Seino, H. Nakai, J. Comput. Chem. 38 (2017) 2520.

23

[30] M. S. Gordon, D. G. Fedorov, S. R. Pruitt, L. V. Slipchenko, Chem. Rev. 112 (2012) 632. [31] H. Nakai, Chem. Phys. Lett. 363 (2002) 73. [32] T. Matsubara, Theor. Phys. 14 (1955) 351. [33] D. J. Thouless, Phys. Rev. 107 (1957) 1162. [34] G. Sanyal, S. H. Mandai, S. Guha, D. Mukherjee, Phys. Rev. E 48 (1993) 3373. [35] M. Kobayashi, T. Taketsugu, Theor. Chem. Acc. 134 (2015) 107. [36] X. He, S. Ryu, S. Hirata, J. Chem. Phys. 140 (2014) 024702. [37] S. Hirata, X. He, J. Chem Phys. 138 (2013) 204112. [38] M. W. Schmidt, K. K. Baldridge, J. A. Boatz, S. T. Elbert, M. S. Gordon, J. H. Jensen, S. Koseki, N. Matsunaga, K. A. Nguyen, S. Su, T. L. Windus, M. Dupuis, J. A. Montgomery Jr., J. Comput. Chem. 14 (1993) 1347. [39] M. Kobayashi, H. Nakai, Int. Quant. Chem. 109 (2009) 2227. [40] W. J. Hehre, R. Ditchfield, J. A. Pople, J. Chem. Phys. 56 (1972) 2257. [41] J. D. Dill, J. A. Pople, J. Chem. Phys. 62 (1975) 2921.

24

Figure Captions Figure 1. Schematics of (a) non-integer-type and (b) integer-type occupations in each subsystem

Figure 2. Labeling of central regions of the BA and U polyene chains C40H42

Figure 3. Buffer-size (nb) dependence of the conventional and FON-type DC-CCSD energy errors from the standard CCSD results of BA and U polyene chains C40H42

Figure 4. Conventional and FON-type DC-CCSD energies errors from standard CCSD energy of BA and U polyene chains C40H42 versus CPU times

Figure 5. Buffer-radius (rb) dependence of the conventional and FON-type DC-CCSD energy errors from the standard CCSD calculations of semiconducting (3,2) and metallic (3,3) SWCNTs

25

Table 1. Comparison between conventional and FON-type DC-CCSD methods based on the canonical HF reference

DC-CCSD Occupied orbital {i,j,…} Virtual orbital {a,b,…} 1 0

FON-DC-CCSD Arbitrary orbital {p,q,…}

Excitation operator

ˆi ,a  aa†ai ˆij ,ab  aa†ab†a j ai

ˆp ,q  aq†a p ˆpq ,rs  ar†as†aq a p

Anti-commutation relation

ai† , a j   a j , ai†    ij   † † aa , ab    ab , aa     ab

 a†p , aq   f p pq ,  a p , aq†   f p pq 

ai† a j   ij

a †p aq  f p pq

Orbital Occupation

Contraction CCSD energy of subsystem

S   Ecor r 

f p  1  f   F   p 

aa ab†   ab

C     S  

f p  f   F   p 

 i









, j a ,b

* i



 j a b 2ti j ,a b  ti j ,b a

26



EcSorr  

a p aq†  f p pq

   S  

 p









,q ,r , s

Cp*  q r  s

 2t

 p q ,r s

 t p q ,s r



Table 2. Numbers of electrons of the localization and central regions calculated based on conventional and FON-type DC-CCSD calculations of BA and U polyene chains C40H42

DC-CCSD Central

BA

FON-DC-CCSD U

region

N eL 

N eS  

N eL 

#1, #20

44.000

15.001

44.000

#2, #19

58.000

13.976

#3, #18

72.000

#4, #17

BA

N eS  

U

N eL 

N eS  

N eL 

N eS  

15.282

44.000

15.000

44.000

14.888

58.000

13.946

58.000

13.977

58.000

13.845

13.983

72.000

14.026

72.001

13.983

72.000

13.882

72.000

13.975

72.000

14.104

72.064

13.987

72.041

13.891

#5, #16

72.000

13.974

72.000

14.161

72.129

13.998

73.169

14.025

#6, #15

72.000

13.973

72.000

14.188

72.171

14.005

73.848

14.090

#7, #14

72.000

13.973

72.000

14.200

72.195

14.010

73.935

14.095

#8, #13

72.000

13.973

72.000

14.205

72.208

14.012

73.956

14.095

#9, #12

72.000

13.973

72.000

14.207

72.215

14.013

73.963

14.095

#10, #11

72.000

13.973

72.000

14.208

72.218

14.014

73.965

14.095

Sum

-

281.548

-

285.054

-

282.000

27

-

282.000

Table 3. Buffer-size (nb) dependence of the numbers of (quasi-)occupied orbitals and (quasi-)virtual orbitals for conventional and FON-type DC-CCSD calculations of BA and U polyene chains C40H42. The numbers of CCSD iterations for the determination of excitation amplitudes, Niter, in conventional and FT-type DC-CCSD calculations are also listed.

nb

DC-CCSD

FON-DC-CCSD

BA  N occ

 N vir

U  Niter

 N occ

 N vir

BA  Niter

 N qocc

 N qvir

U  Niter

 N qocc

 N qvir

 Niter

1

16

98

53

16

98

183

17

99

52

19

101

92

2

26

164

42

26

164

163

27

165

42

29

167

41

3

36

230

41

36

230

112

36

230

42

39

233

42

4

46

296

41

46

296

75

46

296

41

49

299

41

5

56

362

41

56

362

74

56

362

41

59

365

42

28

Figure 1. Schematics of (a) non-integer-type and (b) integer-type occupations in each subsystem

29

Figure 2. Labeling of central regions of the BA and U polyene chains C40H42

30

Figure 3. Buffer-size (nb) dependence of the conventional and FON-type DC-CCSD energy errors from the standard CCSD results of BA and U polyene chains C40H42

31

Figure 4. Conventional and FON-type DC-CCSD energies errors from standard CCSD energy of BA and U polyene chains C40H42 versus CPU times

32

Figure 5. Buffer-radius (rb) dependence of the conventional and FON-type DC-CCSD energy errors from the standard CCSD calculations of semiconducting (3,2) and metallic (3,3) SWCNTs

33

34

Highlights 

The fractional-occupation-number based divide-and-conquer coupled-cluster (FT-DC-CC) is proposed.



The FON formalism solves the inconsistency of occupations for the DC-CC method.



The FON-DC-CC method reduces the energy errors from the DC-CC method.



FON-DC-CC reduces the computational costs by adopting smaller buffer sizes.

35