dmrg in quantum chemistry: from its relation to traditional methods … · 2016. 11. 14. · the...

DMRG in Quantum Chemistry:

From its relation to traditional methods to n-orbital

density matrices and beyond

Markus Reiher

Laboratorium für Physikalische Chemie, ETH Zürich, Switzerland

http://www.reiher.ethz.ch

April 2016

DMRG in Quantum Chemistry Markus Reiher 1 / 77

Outline


Non-Relativistic Many-Electron Hamiltonian

many-electron Hamiltonian in position space (Hartree atomic units)

Hel =

N∑i

(−1

2∇2i −

∑I

ZIriI

)+

N∑i

Standard Procedure: Construction of Many-Electron Basis

Construct many-electron (determinantal) basis set {ΦI} from a given(finite) one-electron (orbital) basis set φi

From the solution of the Roothaan–Hall equations, one obtains n

orbitals from n one-electron basis functions.

From the N orbitals with the lowest energy, the Hartree–Fock (HF)

Slater determinant is constructed.

The other determinants (configurations) are obtained by subsequent

substitution of orbitals in the HF Slater determinant Φ0:

Φ0 → {Φai } → {Φabij } → {Φabcijk} → ... (3)

Determinants are classified by number of ’excitations’

(= substitutions in HF reference determinant).


Standard Full Configuration Interaction (FCI)

Including all possible excited Slater determinants for a finite or infinite

one-electron basis set leads to the so-called full CI approach.

Number of Slater determinants nSD for N spin orbitals chosen from a

set of L spin orbitals (colloquial: N electrons in L spin orbitals):

nSD =

(L

N

)=

L!

N !(L−N)! (4)

Example: There are ≈ 1012 different possibilities to distribute 21electrons in 43 spin orbitals.

CI calculations employing a complete set of Slater determinants are

only feasible for small basis sets.


Truncated CI Wave Functions

Standard recipe to avoid the factorial scaling of the many-electron

basis-set size: truncate basis! Note: basis is pre-defined!

Assumption: Substitution hierarchy is a useful measure to generate a

systematically improvable basis set.

CIS: all singly-(S)-excited determinants are included:

ΨCISel = C0Φ0 +∑(ai)

C(ai)Φai (5)

CISD: all singly- and doubly-(D)-excited determinants are included:

ΨCISDel = C0Φ0 +∑(ai)

C(ai)Φai +

∑(ai)(bj)

C(ai,bj)Φabij (6)

C0, C(ai), C(ai,bj) ∈ {CI} (7)DMRG in Quantum Chemistry Markus Reiher 6 / 77

Determination of the CI Expansion Coefficients CI

ECIel =

〈ΨCIel

∣∣Hel ∣∣ΨCIel 〉〈ΨCIel

∣∣ΨCIel 〉 =∑

I,J C∗ICJ 〈ΦI |Hel |ΦJ〉∑

I,J C∗ICJ 〈ΦI |ΦJ〉

(8)

CI determined from variational principle:

Calculate all derivatives ∂ECIel / ∂C∗K and set them equal to zero, which

yields the CI eigenvalue problem:

H ·C = Eel ·C (9)

Essential: H is constructed from matrix elements 〈ΦI |Hel |ΦJ〉 inpre-defined determinantal basis {ΦI}


Standard ’Technical’ Trick: Second Quantization

Operators and wave functions are expressed in terms of creation and

annihilation operators to implement the Slater–Condon rules for the

evaluation of matrix elements 〈ΦI |Hel |ΦJ〉 directly into the formalism.

Hel in second quantization (i, j, k, l are spin orbital indices):

⇒ Hel =∑ij

〈φi|h(i) |φj〉 c†i cj

+1

2

∑ijkl

〈φi(1) 〈φk(2)| g(1, 2) |φl(2)〉φj(1)〉c†i c†jckcl (10)

CI wave function in second quantization:

ΨFCIel = C0Φ0 +∑(ai)

C(ai)c†aciΦ0 +

∑(ai)(bj)

C(ai,bj)c†bcjc

†aciΦ0 · · · (11)


CI Energy in Second Quantization

ECIel =D

ΨCIel

˛̨̨Hel

˛̨̨ΨCIel

E(12)

=

LXij

nSDXIJ

C∗ICJ tIJij| {z }

γij

〈φi(1)|h(1) |φj(1)〉| {z }≡hij

+LXijkl

nSDXIJ

C∗ICJTIJijkl| {z }

Γijkl

〈φi(1) 〈φk(2)| g(1, 2) |φl(2)〉φj(1)〉| {z }gijkl

(13)

=LXij

γijhij +LXijkl

Γijklgijkl (14)

tIJij or TIJijkl are matrix elements of determinantal basis functions over

pairs or quadruples of elementary operators c† and c.

γij are Γijkl are density matrix elements (compare: RDMFT and 2RDM

methods).


Is there a better way to construct the finite-dimensional

determinantal basis set in order to avoid the factorial scaling?


Coupled-Cluster — An Advanced CI-type Wave FunctionAnsatz:

ΨCCel = exp (T ) Φ0 (15)

Excitation operator:

T = T1 + T2 + T3 + · · · (16)

where

Tα =X

ab · · ·| {z }α times

ij · · ·| {z }α times

cluster-amplitudesz}|{tab···ij··· · · · c†bcjc

†aci| {z }

α pairs c†c

⇒ T1 =Xai

tai c†aci (17)

Notation:

CCS (T = T1), CCSD (T = T1 + T2), CCSDT (T = T1 + T2 + T3) ,· · ·Coupled-cluster improves on truncated CI, because certain (disconnected)

higher excited configurations (e.g., tai c†acit

bcjkc†cckc

†bcj) are included.


However, all of this is too expensive and often too limited

(MRCC would be nice, but requires 5e-RDMs ...)

DFT is the workhorse in transition metal chemistry

(structures(!), energies, spectroscopic signatures)

in critical cases, its accuracy is difficult to assess

... illustrated by an example


The Spin State Problem in Quantum ChemistryTo predict the correct energy gaps between states of different spin S

is crucial in chemistry (e.g., two-state reactivity); in kJ/mol

Present-day DFT has difficulties to cope with this problem

E.g., the results strongly depend on exact exchange admixture c3

Example: the thermal spin-crossover complex [Fe(phen)2(NCS)2]

N

N

N

N

Fe

N

NCS

CS

0 0.05 0.1 0.15 0.2

c3 parameter

−75

−25

25

75

∆ E

a/b / k

J/m

ol

S=0/S=2 splitting

S=0/S=1 splitting

M. Reiher, Inorg. Chem. 41 (2002) 6928;

M. Reiher, Faraday Discus. 135 (2007) 97

DFT ∆EeLS/HS

B3LYP(c3=0.00) 82.6

BP86/RI 66.0

BLYP 48.1

B3LYP(c3=0.12) 11.0

B3LYP(c3=0.13) 5.5

B3LYP(c3=0.14) −0.3B3LYP(c3=0.15) −6.1B3LYP −33.4HF −407.5DMRG in Quantum Chemistry Markus Reiher 13 / 77

Ab initio Methods for benchmarks needed

Well-established for open-shell transition metal complexes:

CASSCF/CASPT2 and MRCI

(incl. RASSCF/RASPT2)


Orbital subspaces in CAS approaches

inactive orbitals ~ 50 - 100

active orbitals CASSCF: up to 18

DMRG-SCF: up to 100

virtual orbitals ~ 102 - 103

orbitals are subdivided in doubly occupied

inactive orbitals, active orbitals and

unoccupied virtual orbitals

huge virtual orbital subspace is not

considered in CAS approaches (but should

be (dynamic correlation)!)

Exponential scaling of many-particle basis

set size permits exact diagonalization up to

only CAS(18,18)


Ab initio Methods for benchmarks needed

Well-established for open-shell transition metal complexes:

CASSCF/CASPT2 and MRCI

(incl. RASSCF/RASPT2)

... but what if active orbital space is too large?

Promising new approaches?


Parameterization of the Wave Function

|Ψ〉 =∑

i1i2···iN

Ci1i2···iN |i1〉 ⊗ |i2〉 ⊗ · · · ⊗ |iN 〉

Restricted sum over basis states with a certain excitation pattern

→ yields a pre-defined (!) many-particle basis setConfiguration Interaction ansatz

|CI〉 =(

1 +∑µ

Cµτ̂µ

)|HF〉

Coupled Cluster ansatz

|CC〉 =(∏

µ

[1 + tµτ̂µ])|HF〉

numerous specialized selection/restriction protocols


Instead of standard CI-type calculations by

diagonalization/projection

|Ψ〉 =∑

i1i2···iN

Ci1i2···iN |i1〉 ⊗ |i2〉 ⊗ · · · ⊗ |iN 〉

construct CI coefficients from correlations among orbitals

|Ψ〉 =∑

i1i2···iN

Ci1i2···iN |i1〉 ⊗ |i2〉 ⊗ · · · ⊗ |iN 〉


Density Matrix Renormalization Group: Principles

New wavefunction parametrization:

|ψ〉 =∑σ

cσ|σ〉 −→ |ψ〉 =∑σ

Mσ1 Mσ2 · · · MσL |σ〉

complete active space (CAS) partitioning and iterative sweeping:

In essence: dimension reduction by least-squares fitting


Reviews on DMRG in Quantum Chemistry

1 Ö Legeza, R. M. Noack, J. Sólyom and L. Tincani, Applications of Quantum Information in the Density-Matrix

Renormalization Group, Lect. Notes Phys., 739,653-664 (2008)

2 G. K.-L. Chan, J. J. Dorando, D. Ghosh, J. Hachmann, E. Neuscamman, H. Wang, T. Yanai, An Introduction to the

Density Matrix Renormalization Group Ansatz in Quantum Chemistry, Prog. Theor. Chem. and Phys., 18, 49 (2008)

3 D. Zgid and G. K.-L. Chan, The Density Matrix Renormalisation Group in Quantum Chemistry, Ann. Rep. Comp.

Chem., 5, 149, (2009)

4 K. H. Marti, M. Reiher, The Density Matrix Renormalization Group Algorithm in Quantum Chemistry, Z. Phys. Chem.,

224, 583-599 (2010)

5 G. K.-L. Chan and S. Sharma, The density matrix renormalization group in quantum chemistry, Ann. Rev. Phys. Chem.,

62, 465 (2011)

6 K. H. Marti, M. Reiher, New Electron Correlation Theories for Transition Metal Chemistry, Phys. Chem. Chem. Phys.,

13, 6750-6759 (2011)

7 Y. Kurashige, Multireference electron correlation methods with density matrix renormalisation group reference functions,

Mol. Phys. 112, 1485-1494 (2014)

8 S. Wouters and D. Van Neck, The density matrix renormalization group for ab initio quantum chemistry, Eur. Phys. J.

D 68, 272 (2014)

9 T. Yanai, Y. Kurashige, W. Mizukami, J. Chalupský, T. N. Lan, M. Saitow, Density matrix renormalization group for ab

initio Calculations and associated dynamic correlation methods,Int. J. Quantum Chem. 115, 283-299 (2015)

10 S. Szalay, M. Pfeffer, V. Murg, G. Barcza, F. Verstraete, R. Schneider, Ö. Legeza, Tensor product methods and

entanglement optimization for ab initio quantum chemistry, arXiv:1412.5829 (2015)


DMRG Parameters

Active space selection, as in MCSCF, but need additional parameters:

Number of renormalized basis states m

Orbital ordering

Choice of initial environment guess

These parameters do depend on each other.

S. Keller, M. Reiher, Chimia, 2014, 68, 200–203


Two variants of DMRG:

Traditional DMRG

|ψ〉 =∑LR CLR|σL〉 ⊗ |σR〉

coefficients valid for one

bipartition into L and R

(need basis

transformations)

Excited states only with

state averaging

faster for ground states?

MPS DMRG (2nd generation)

|ψ〉 =∑σM

σ1 Mσ2 · · · MσL |σ〉

coefficients valid for the whole

system

Easy and efficient

implementation of

〈φ|ψ〉〈ψ|a†i Ĥaj |ψ〉3- and 4-body RDMs

excited states


Matrix product operators (MPO)

Ĥ =L∑ij σ

tij ĉ†iσ ĉjσ +

1

2

L∑ijklσ σ′

Vijklĉ†iσ ĉ†kσ′ ĉlσ′ ĉjσ,


MPO compression

Consider the terms c†1c†2c5c7 + c

†1c†3c6c7


MPS-MPO operations: expectation values


Spin symmetry adaptation: MPS and MPO put together

F[a′]i′j′(l + 1) =

∑aijkss′

Wigner-9j[jsj′

aka′is′i′

]M

[s′]†

i′i W[k]ss

′

aa′ F[a]ij (l)M

[s]j′j

S. Keller, M. Reiher, J. Chem. Phys. 144, 134101 (2016)


The QCMaquis DMRG program

Input: tij and Vijkl integrals, target state quantum numbers

The system (atom, molecule, ...) is encoded in tij

Co-developed with the ALPS MPS project (Troyer).

Integrated into the Molcas program package.

Source code available from www.reiher.ethz.ch/software

Parallelized by MPI and threading

C++ / C++11, 60k lines core program + 50k lines associated libs

+ standard libs (boost, etc.)

S. Keller, M. Dolfi, M. Troyer, M. Reiher, J. Chem. Phys. 143, 244118 (2015)


www.reiher.ethz.ch/software

A well working example: FCI for N2 (14e, 28o)


Does DMRG Work for Compact Molecules?

Original ’opinion’ in the DMRG community:

Works only for pseudo-one-dimensional, non-compact systems!

Test for a mononuclear transition metal system CAS(10,10): CoH

m Esinglet/Eh Etriplet/Eh ∆E/kJmol−1

64 −1381.952 054 −1381.995 106 113.0376 −1381.952 063 −1381.995 109 113.0291 −1381.952 070 −1381.995 110 113.00109 −1381.952 073 −1381.995 110 112.99CAS(10,10) −1381.952 074 −1381.995 110 112.99CASPT2(10,10) −1382.189 527 −1382.241 333 130.57DFT/BP86 −1383.504 019 −1383.585 212 213.1DFT/B3LYP −1383.202 267 −1383.279 574 203.0

K. Marti, I. Malkin Ondik, G. Moritz, M. Reiher, J. Chem. Phys 128 (2008) 014104


Example: Analysis of Spin Density Distributions

with DMRG


Noninnocent Iron Nitrosyl Complexes

(a) Fe(salen)(NO) conformation a

(b) Fe(salen)(NO) conformation b

(c) Fe(porphyrin)(NO)

transition metal nitrosyl complexes

have a complicated electronic

structure

qualitatively different spin

densities reported by Conradie

& GhoshJ. Conradie, A. Ghosh, J. Phys. Chem. B 2007, 111,

12621.

systematic comparison of DFT

spin densities with CASSCF:K. Boguslawski, C. R. Jacob, M. Reiher, J. Chem.

Theory Comput. 2011, 7, 2740;

see also work by K. Pierloot et al.


DFT Spin Densities: A Case Study

(a) OLYP (b) OPBE (c) BP86 (d) BLYP

(e) TPSS (f) TPSSh (g) M06-L (h) B3LYP

Only for high-spin complexes similar spin densities are obtained

⇒ [Fe(NO)]2+ moiety determines the spin density


The Model System for Accurate Reference Calculations

zyx

Fe

N

O

�

dpc dpc

dpcdpc

Structure:

Four point charges of −0.5 e model a square-planarligand field (ddp = 1.131 Å)

⇒ Similar differences in DFT spin densities as present forlarger iron nitrosyl complexes

Advantage of the small system size:

Standard correlation methods (CASSCF, . . .) can be efficiently

employed

Study convergence of the spin density w.r.t. the size of the active

orbital space

K. Boguslawski, C. R. Jacob, M. Reiher, J. Chem. Theory Comput. 2011, 7, 2740.


DFT Spin Densities

(a) OLYP (b) OPBE (c) BP86 (d) BLYP

(e) TPSS (f) TPSSh (g) M06-L (h) B3LYP

Spin density isosurface plots Spin density difference plots w.r.t.

OLYP

⇒ Similar differences as found for the large iron nitrosyl complexes


Reference Spin Densities

from Standard Electron Correlation Methods


Reference Spin Densities from CASSCF Calculations

Defining the active orbital space:

Minimal active space: Fe 3d- and both NO π∗-orbitals ⇒ CAS(7,7)

Consider also both NO π-orbitals ⇒ CAS(11,9)

Additional shell of Fe d-orbitals (double-shell orbitals) is gradually

included ⇒ CAS(11,11) to CAS(11,14)

Include one ligand σ-orbital and the antibonding σ∗-orbital upon the

CAS(11,11) ⇒ CAS(13,13)

⇒ Analyze convergence of spin density w.r.t. the dimension of the activeorbital space


CASSCF Results: Oscillating Spin Densities

CAS(7,7) CAS(11,9) CAS(11,11) CAS(11,12) CAS(11,14)

CAS(11,13) CAS(13,13) CAS(13,14) CAS(13,15) CAS(13,16)

CAS(11,14): spin

density; all others:

spin density

differences

stable CAS with all

important orbitals is

difficult to obtain

⇒ Reference spindensities for very

large CAS required


DMRG Spin Densities — Measures of Convergence

Qualitative convergence measure: spin density difference plots

Quantitative convergence measure:

∆abs =

∫|ρs1(r)− ρs2(r)|dr < 0.005 (18)

∆sq =

√∫|ρs1(r)− ρs2(r)|2dr < 0.001 (19)

Quantitative convergence measure: quantum fidelity Fm1,m2

Fm1,m2 = |〈Ψ(m1)|Ψ(m2)〉|2 (20)

⇒ Reconstructed CI expansion of DMRG wave function can be used!K. Boguslawski, K. H. Marti, M. Reiher, J. Chem. Phys. 2011, 134, 224101.


DMRG Spin Densities for Large Active Spaces∆abs and ∆sq for DMRG(13,y)[m] calculations w.r.t. DMRG(13,29)[2048] reference

Method ∆abs ∆sq Method ∆abs ∆sq

DMRG(13,20)[128] 0.030642 0.008660 DMRG(13,29)[128] 0.032171 0.010677

DMRG(13,20)[256] 0.020088 0.004930 DMRG(13,29)[256] 0.026005 0.006790

DMRG(13,20)[512] 0.016415 0.003564 DMRG(13,29)[512] 0.010826 0.003406

DMRG(13,20)[1024] 0.015028 0.003162 DMRG(13,29)[1024] 0.003381 0.000975

DMRG(13,20)[2048] 0.014528 0.003028

DMRG(13,20;128) DMRG(13,20;256) DMRG(13,20;512) DMRG(13,20;1024) DMRG(13,20;2048)

DMRG(13,29;128) DMRG(13,29;256) DMRG(13,29;512) DMRG(13,29;1024) DMRG(13,29;2048)


Importance of Empty Ligand Orbitals

Table: Some important Slater determinants with large CI weights from DMRG(13,29)[m]Upper part: Slater determinants containing an occupied dx2−y2 -double-shell orbital (marked in

bold face). Bottom part: Configurations with occupied ligand orbitals (marked in bold face). 2:

doubly occupied; a: α-electron; b: β-electron; 0: empty.

CI weight

Slater determinant m = 128 m = 1024

b2b222a0a0000000 0000000 a 00000 0.003 252 0.003 991

bb2222aa00000000 0000000 a 00000 −0.003 226 −0.003 611222220ab00000000 0000000 a 00000 −0.002 762 −0.003 328ba2222ab00000000 0000000 a 00000 0.002 573 0.003 022

b2a222a0b0000000 0000000 a 00000 −0.002 487 −0.003 017202222ab00000000 0000000 a 00000 0.002 405 0.002 716

b222a2a0b0000000 0000000 0 0000a 0.010 360 0.011 558

22b2a2a0a0000000 0000000 0 b0000 0.009 849 0.011 366

22b2a2a0b0000000 0000000 0 a0000 −0.009 532 −0.011 457b2222aab00000000 0000000 0 0000a −0.009 490 −0.010 991a2222baa00000000 0000000 0 0000b −0.009 014 −0.010 017b2b222a0a0000000 0000000 0 0a000 0.008 820 0.010 327


Assessment of CASSCF Spin Densities (Differences)

CAS(11,11) CAS(11,12) CAS(11,13) CAS(11,14)

CAS(13,13) CAS(13,14) CAS(13,15) CAS(13,16)

⇒ CASSCF spin densities oscillate around DMRG reference distribution


DFT–DMRG Spin Densities Differences

OLYP OPBE BP86 BLYP

TPSS TPSSh M06-L B3LYP

K. Boguslawski, K. H. Marti, Ö. Legeza, M. Reiher, J. Chem. Theory Comput. 8 2012 1970


Analyzing DMRG and correlated wave functions

with concepts from quantum information theory


Entanglement Measures for Embedded Subsystems

See pioneering work by Ö. Legeza within QC-DMRG

Consider one or two orbitals embedded in a CAS

Measure for entanglement of states on orbital i with those defined on

environment:

Ö. Legeza, J. Sólyom, Phys. Rev. B 2003, 68, 195116.

von-Neumann-type single-orbital entropy

s(1)i = −4∑

α=1

ωα,i lnωα,i (21)

(ωα,i is an eigenvalue of the 1o-RDM of spatial orbital i — states

defined on all other orbitals of the CAS have been traced out)



Measure for entanglement of states living on orbital i and orbital j

with those in the environment:

von-Neumann-type two-orbital entropy

s(2)ij = −16∑α=1

ωα,ij lnωα,ij (22)

(ωα,ij is an eigenvalue of the 2o-RDM of two spatial orbitals i and j

— states defined on all other orbitals of the CAS have been traced

out)

Note the difference between ne-RDMs and no-RDMs !



s(2)ij contains also the ’on-site’ entropies for the two orbitals

⇒ Subtract these contributions to obtain the ’inter-orbital entropy’:J. Rissler, R.M. Noack, S.R. White, Chem. Phys. 2006, 323, 519.

Mutual information

Iij ∝ s(2)ij − s(1)i − s(1)j (23)

Successfully applied to optimize orbital ordering and to enhance

DMRG convergence by Ö. Legeza


How To Read the (New) Mutual Information Plots


Entanglement and Orbitals — Single Orbital Entropy

0 5 10 15 20 25 300

0.1

0.2

0.3

0.4

0.5

0.6

0.7

Orbital index

s(1

)

K. Boguslawski, P. Tecmer, Ö. Legeza, M. Reiher,

J. Phys. Chem. Lett. 2012, 3, 3129.

Three groups of orbitals

⇒ large single orbital entropy⇒ medium single orbital entropy⇒ tiny single orbital entropyConfigurations belonging to the

third block have small CI ⇒important for dynamic

correlation

Note: more structure than in

spectrum of 1e-RDM

zyx

Fe

N

O

�

dpc dpc

dpcdpc

4 point charges in xy-plane at dpc = 1.133 Å

Natural orbital basis: CAS(11,14)SCF/cc-pVTZ

DMRG(13,29) with DBSS (mmin = 128,mmax = 1024)


Entanglement and Orbitals — Mutual Information

1

3

5

7

9

11

1315

17

19

21

23

25

27

29

2

4

6

8

10

12

1416

18

20

22

24

26

28

A

100

10−1

10−2

10−3

K. Boguslawski, P. Tecmer, Ö. Legeza, M. Reiher,

J. Phys. Chem. Lett. 2012, 3, 3129.

Three groups of orbitals

⇒ high entanglement⇒ medium entanglement⇒ weak entanglementStrong entanglement for pairs:

(d, π∗)⇐⇒ (d, π∗)∗π ⇐⇒ π∗

σMetal ⇐⇒ σLigand⇒ Important for static correlation (⇐⇒

chemical intuition of constructing a

CAS)

zyx

Fe

N

O

�

dpc dpc

dpcdpc

4 point charges in xy-plane at dpc = 1.133 Å

Natural orbital basis: CAS(11,14)SCF/cc-pVTZ

DMRG(13,29) with DBSS (mmin = 128,mmax = 1024)


Entanglement Measures can Monitor

Bond Breaking/Formation Processes: Dinitrogen

K. Boguslawski, P. Tecmer, G. Barcza, O. Legeza, M. Reiher, J. Chem. Theory Comput. 9 2013

2959–2973 [arxiv: 1303.7207]


Bond Breaking in Dinitrogen at 1.12 Ångström

Part of entanglement is already encoded in molecular orbitals changing with the structure !

Atomic-like non-orthogonal basis fcts. exhibit large entanglement measures among each other.


Artifacts of Small Active Space CalculationsComparison of DMRG(11,9)[220] and DMRG(11,11)[790] to

DMRG(13,29)[128,1024,10−5] for [Fe(NO)]2+

1

3

5

7

9

2

4

6

8

A

100

10−1

10−2

10−3

1

3

5

7

9

11

2

4

6

8

10

A

100

10−1

10−2

10−3

1

3

5

7

9

11

1315

17

19

21

23

25

27

29

2

4

6

8

10

12

1416

18

20

22

24

26

28

A

100

10−1

10−2

10−3

Iij (and s(1)i) overestimated for

small active spaces

⇒ Entanglement of orbitals too large⇒ Missing dynamic correlation is

captured in an artificial way


How to choose an active space?


How to choose an active space? — Opinions

“... the choice of the active space actually used in more complex

systems is highly subjective and can lead to serious problems.“

R. D. Bach in The Chemistry of Peroxides, Z. Rappoport (Ed.), John Wiley & Sons, (2006) 4.

”CAS-based methods are another alternative, although the selection

of the active space is a tremendous challenge.“

Y. Shao, L. Molnar, Y. Jung, J. Kussmann, C. Ochsenfeld et al., Phys. Chem. Chem. Phys., 8, (2006) 3172–3191.

“Choosing the “correct” active space for a specific application is by

no means trivial; often the practitioner must “experiment” with

different choices in order to assess adequacy and convergence

behavior. While every chemical system poses its own challenges,

certain rules of thumb apply.”

P. Å. Malmqvist, K. Pierloot, A. R. M. Shahi, C. J. Cramer, L. Gagliardi, J. Chem. Phys., 128, (2008) 204109.


How to choose an active space? — Recipes

B. O. Roos (1989)The Complete Active Space Self-Consistent Field Method and Its Applications in

Electronic Structure Calculations; in Ab Initio Methods in Quantum Chemistry,

K.P. Lawley (Ed.), John Wiley & Sons Ltd., 399–446.

- most active orbitals should appear paired (one highly occupied and

one corresponding almost empty orbital)

- conjugated and aromatic bonds should be included in the CAS

- both the bonding and antibonding orbitals of a bond that is broken

have to be included

- orbitals describing C-H bonds are not to be included in active space


How to choose an active space? — Recipes

J. M. Bofill, P. Pulay (1989)J. Chem. Phys. 90, 3637–3646.

- unrestricted Hartree-Fock natural orbitals (UNOs) are used as

starting orbitals for CASSCF calculations

- orbitals with occupation numbers between 0.02 and 1.98 are

selected for the active space

- the idea has been explored further by comparison to DMRG results:

S. Keller, K. Boguslawski, T. Janowski, M. Reiher, P. Pulay (2015)J. Chem. Phys. 142, 244104.


Can one choose the CAS in an automated way?


Exploit Two Advantages of DMRG for the CAS choice

DMRG results obtained after (micro)iterations, qualitative structure

of wave function converges quickly

The CAS is not limited by exponentially scaling basis size (the scaling

depends on the no. of orbitals, not on the no. of electrons; 100

orbitals are about the limit)


Example: Transition Structure of Ozone

0 5 10 15 20 25 30DMRG microiteration steps

-224.40

-224.38

-224.35

-224.33

-224.30

-224.28

ener

gy

/ H

artr

ee

HFCASCIDMRG

O3

transition state energy

15 20 25 30iteration

-224.3845

-224.3840

-224.3835

-224.3830

ener

gy

/ H

artr

ee


0.0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

abs.

CI

coef

fici

ent

SD1: 111001000000110110SD2: 110110000000111001SD3: 111100000000111100SD4: 111100000000111001SD5: 111010000000110101SD6: 101101000000111100

O3

transition state CI coefficients

G. Moritz, M. Reiher, J. Chem. Phys. 126 (2007) 244109

(see this reference also for a ’first-generation’ DMRG flow chart)




-224.40

-224.38

-224.35

-224.33

-224.30

-224.28

ener

gy

/ H

artr

ee

HFCASCIDMRG

O3



-224.3845

-224.3840

-224.3835

-224.3830

ener

gy

/ H

artr

ee


0.0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

abs.

CI

coef

fici

ent

SD1: 111001000000110110SD2: 110110000000111001SD3: 111100000000111100SD4: 111100000000111001SD5: 111010000000110101SD6: 101101000000111100

O3





Exploit Two Advantages of DMRG for the CAS choice

DMRG is iterative

DMRG can handle large CAS sizes

=⇒ ... towards an automated CAS determination


Partially converged entanglement information is sufficient

E = -2658.144420 EH E = -2665.877397 EH E = -2665.878045 EH

E = -1449.266845 EH E = -1449.290301 EH E = -1449.302426 EH

HF orbitals, m = 2004 sweepsHF guess

HF orbitals, m = 2004 sweepsHF guess

HF orbitals, m = 5008 sweeps

CI-DEAS guess


CI-DEAS guess


CI−DEAS guess


CI−DEAS guess

sufficiently accurate entanglement information can be obtained from partially

converged calculations C. J. Stein, M. Reiher, J. Chem. Theory Comput, 2016, 12, 1760–1771.


Complete Automation of CAS Selection

generate initial orbitals recommended: CASSCF with small CAS or DMRG-SCF

initial DMRG calculation with a large CAS around the Fermi level

recommended settings: CI-DEAS guess, m = 500, 8 sweeps

max. si (1) > 0.14 ?

single-configuration case

generate threshold diagrams

calculate entanglement measures

plateaus?Can orbitals with

si (1) < 1-2 % max. si (1) be excluded? include orbitals kept in the first,

clearly identifiable plateau at low thresholds

converge calculation with

CASSCF or DMRG-SCF

consistency test: Do entanglement measures of the final calculation agree qualitatively with those of the initial calculation?YES

YES

NO

NO

Flowchart for the Automated Selection of Active Orbital Spaces

select CAS anyway

YES

Is it feasible to converge a calculation with this CAS size?

YES

NO

consistency test:

increase initial CAS

NO

STOP: unfeasible for DMRG

C. J. Stein, M. Reiher, J. Chem. Theory Comput, 2016, 12, 1760–1771.


Same CAS selection for different types of orbitals

MnO4-

HF orbitals

CAS(10,10)-SCF orbitals

split-localized orbitals

DMRG(38,25)[500]-SCF orbitals

the same CAS is automatically selected for all orbital bases

considered C. J. Stein, M. Reiher, J. Chem. Theory Comput, 2016, 12, 1760–1771.DMRG in Quantum Chemistry Markus Reiher 71 / 77

Subtile correlation effects are automatically captured

DMRG-SCF orbitalsm = 1000, 20 sweeps

HF guess

CrF63- CrF6

correlation effects attributed to the covalency of the bonds are automatically

accounted for

C. J. Stein, M. Reiher, J. Chem. Theory Comput, 2016, 12, 1760–1771.


The Cu2O2+2 torture track

·−1

bis(µ-oxo) peroxo

F

CR-CCSD(TQ)LCAS(16,14)-SCFDMRG(48,36)[500]-SCFDMRG(32,28)[1000]-SCF#DMRG(48,36)[500]-SCF

automated approach gives qualitatively correct wave

function C. J. Stein, M. Reiher, J. Chem. Theory Comput, 2016, 12, 1760–1771.

discrepancy to coupled-cluster result can be explained by (still) missing

dynamical correlation C. J. Cramer et al., JPC A (2006), 110, 1991.


Regularizing effect of short-range DFT: Water

E. D. Hedeg̊ard, S. Knecht, J. S. Kielberg, H. J. A. Jensen, and M. Reiher, J. Chem. Phys. 142 2015 224108


DMRG–srDFT for the WCCR10 test set

N

NCu N

N

N

NCu N

N

7Å

Calculated dissociation energies in kJ/mol

Method De (kJ/mol) D0 (kJ/mol)

DMRG[2000](30,22) 173.5 165.1

DMRG[2000](20,18) 169.9 161.5

DMRG[2000](10,10) 132.8 124.3

DMRG[2000](30,22)-srPBE 225.1 216.6

DMRG[2000](20,18)-srPBE 227.9 219.4

DMRG[2000](10,10)-srPBE 216.5 208.0

PBE 240.2 231.8

PBE (full complex/def2-TZVP) 257.5 249.0

PBE (full complex/def2-QZVPP from WCCR10) 247.5 239.0

Exp. (from WCCR10) 226.7 218.2

E. D. Hedeg̊ard, S. Knecht, J. S. Kielberg, H. J. A. Jensen, and M. Reiher, J. Chem. Phys. 142 2015 224108;WCCR10: T. Weymuth, E. P. A. Couzijn, P. Chen, M. Reiher, J. Chem. Theory Comput. 10 2014 3092


Conclusion

new and efficient second-generation DMRG program for full

Schrödinger–Coulomb (Dirac–Gaunt/Breit) Hamiltonian available:

QCMaquis

HF-, split-localized- and CASSCF-orbitals allow to select an active

space that is in line with the literature suggestions

the whole CAS evaluation procedure is fast

... and can be automatized

dynamic correlation introduced through short-range DFT regularizes

the CAS


Acknowledgments

Christopher Stein, Sebastian Keller

Stefan Knecht, Arseny Kovyrshin

Yingjin Ma, Erik Hedeg̊ard

Andrea Muolo, Pawel Tecmer

Katharina Boguslawski

... as well as the rest

of the group −→

2nd editionFinancial Support: ETH, SNF, FCI

Thank you for your kind

attention!


dmrg in quantum chemistry: from its relation to traditional methods … · 2016. 11. 14. · the...

Documents