FCIQMC, (Un)linked Stochastic Coupled Cluster Theory and Other
Transcript of FCIQMC, (Un)linked Stochastic Coupled Cluster Theory and Other
FCIQMC,(Un)linked Stochastic Coupled Cluster Theory
and Other Animals
Alex Thom
Department of Chemistry, University of Cambridge
ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
Electronic Structure Theory: Scaling
100 101 102 103
# Water molecules
10-6
10-4
10-2
100
102
104
106
108
1010
1012
1014Calculation Time / s
minute
hour
day
week
yeardecade
centurymillennium
239Pu half life
aeon
age of universe
force-field
poor SCF
good SCF
MP2
CCSD
CCSD(T)FCI
ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
Electronic Structure Theory: Accuracy
HF DFT:GGADFT:Hybrid MP2 CCSD CCSD(T) CCSDT CCSDT(Q)Method
−80
−60
−40
−20
0
20
40
60
80
Devia
tion f
rom
experi
ment
/ kJ
/mol
Helgaker T. et al., Molecular Electronic Structure TheoryCurtiss L.A., Raghavachari K., Redfern P.C., Pople, J.A. J. Chem. Phys. 106, 1063 (1997)Klopper W., Helgaker T. J. Phys. B 32, R103 (1999)Klopper W., Ruscic B., Tew D., Bischoff F., Wolfsegger S. Chem. Phys. 356, 14 (2009)Klopper W., Bachorz R., Hattig C., Tew D. Theor. Chem. Acc. 126. 289 (2010)
ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
Determinant Space
Hartree-Fock Triple excitation
I For an N -electron system, N spin-orbitals are chosen out of2M spin-orbitals {φ1, φ2, ..., φ2M} to form a SlaterDeterminant,|Di〉.
I Complete space of determinants is finite, but exponentiallygrowing in N and M as
(2MN
)
ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
Determinant Space
Hartree-Fock Triple excitation
I For an N -electron system, N spin-orbitals are chosen out of2M spin-orbitals {φ1, φ2, ..., φ2M} to form a SlaterDeterminant,|Di〉.
I Complete space of determinants is finite, but exponentiallygrowing in N and M as
(2MN
)ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
Full Configuration Interaction
I We can project the many-electron Hamiltonian into the spaceof Slater Determinants and solve in this space.Hij = 〈Di|H|Dj〉.
I Solve H|ΨCI〉 = E|ΨCI〉 as |ΨCI〉 =∑
i ci|Di〉.I Iterative diagonalization of the sparse Hamiltonian in this
space gives the “Full Configuration Interation” (FCI) solution.
I As H contains at most 2-electron operators, matrix elementsbetween determinants are a simple combination of one- andtwo-electron Hamiltonian integrals.
I Energy is variationally minimized — all of the basis setcorrelation energy captured.
I Exponentially scaling with system size (N or M)
ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
Full Configuration Interaction
I We can project the many-electron Hamiltonian into the spaceof Slater Determinants and solve in this space.Hij = 〈Di|H|Dj〉.
I Solve H|ΨCI〉 = E|ΨCI〉 as |ΨCI〉 =∑
i ci|Di〉.I Iterative diagonalization of the sparse Hamiltonian in this
space gives the “Full Configuration Interation” (FCI) solution.
I As H contains at most 2-electron operators, matrix elementsbetween determinants are a simple combination of one- andtwo-electron Hamiltonian integrals.
I Energy is variationally minimized — all of the basis setcorrelation energy captured.
I Exponentially scaling with system size (N or M)
ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
Full Configuration Interaction
I We can project the many-electron Hamiltonian into the spaceof Slater Determinants and solve in this space.Hij = 〈Di|H|Dj〉.
I Solve H|ΨCI〉 = E|ΨCI〉 as |ΨCI〉 =∑
i ci|Di〉.I Iterative diagonalization of the sparse Hamiltonian in this
space gives the “Full Configuration Interation” (FCI) solution.
I As H contains at most 2-electron operators, matrix elementsbetween determinants are a simple combination of one- andtwo-electron Hamiltonian integrals.
I Energy is variationally minimized — all of the basis setcorrelation energy captured.
I Exponentially scaling with system size (N or M)
ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
FCI Quantum Monte Carlo
I Solutions to H|Ψ〉 = E|Ψ〉 also solve e−τH |Ψ〉 = e−τE |Ψ〉.
I Propagate ∂|Ψ〉∂τ = −H|Ψ〉.
I Lowest energy eigenfunction |ΨCI〉 becomes dominant.
I Let |ΨCI〉 =∑
i ci|Di〉
∂ci∂τ
= −∑j
Hijcj.
I Represent ci as populations of signed psips (walkers)propagating through space.
George H. Booth, AJWT, and Ali Alavi, J. Chem. Phys. 131, 054106 (2009)
James B. Anderson, J. Chem. Phys. 63, 1499 (1975)
ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
FCI Quantum Monte Carlo
I Solutions to H|Ψ〉 = E|Ψ〉 also solve e−τH |Ψ〉 = e−τE |Ψ〉.I Propagate ∂|Ψ〉
∂τ = −H|Ψ〉.I Lowest energy eigenfunction |ΨCI〉 becomes dominant.
I Let |ΨCI〉 =∑
i ci|Di〉
∂ci∂τ
= −∑j
Hijcj.
I Represent ci as populations of signed psips (walkers)propagating through space.
George H. Booth, AJWT, and Ali Alavi, J. Chem. Phys. 131, 054106 (2009)
James B. Anderson, J. Chem. Phys. 63, 1499 (1975)
ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
FCI Quantum Monte Carlo
I Solutions to H|Ψ〉 = E|Ψ〉 also solve e−τH |Ψ〉 = e−τE |Ψ〉.I Propagate ∂|Ψ〉
∂τ = −H|Ψ〉.I Lowest energy eigenfunction |ΨCI〉 becomes dominant.
I Let |ΨCI〉 =∑
i ci|Di〉
∂ci∂τ
= −∑j
Hijcj.
I Represent ci as populations of signed psips (walkers)propagating through space.
George H. Booth, AJWT, and Ali Alavi, J. Chem. Phys. 131, 054106 (2009)
James B. Anderson, J. Chem. Phys. 63, 1499 (1975)
ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
FCI Quantum Monte Carlo
I Solutions to H|Ψ〉 = E|Ψ〉 also solve e−τH |Ψ〉 = e−τE |Ψ〉.I Propagate ∂|Ψ〉
∂τ = −H|Ψ〉.I Lowest energy eigenfunction |ΨCI〉 becomes dominant.
I Let |ΨCI〉 =∑
i ci|Di〉
∂ci∂τ
= −∑j
Hijcj.
I Represent ci as populations of signed psips (walkers)propagating through space.
George H. Booth, AJWT, and Ali Alavi, J. Chem. Phys. 131, 054106 (2009)
James B. Anderson, J. Chem. Phys. 63, 1499 (1975)
ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
Dynamics
I At time τ , given a psip at j, at time τ + δτ spawn a new oneat randomly chosen i based on −δτHij. This samples
∑j.
I Psips at i die based on .
I Start S = EHF . Once enough psips created change S inresponse to population. e.g. for growth, make S morenegative, so more death.
I Opposite-signed psips at the same determinant annihilate.
ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
Dynamics
∂ci∂τ
= −Hiici −∑j→i
Hijcj.
I At time τ , given a psip at j, at time τ + δτ spawn a new oneat randomly chosen i based on −δτHij. This samples
∑j.
I Psips at i die based on .
I Start S = EHF . Once enough psips created change S inresponse to population. e.g. for growth, make S morenegative, so more death.
I Opposite-signed psips at the same determinant annihilate.
ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
Dynamics
∂ci∂τ
= −Hiici −∑j→i
Hijcj.
I At time τ , given a psip at j, at time τ + δτ spawn a new oneat randomly chosen i based on −δτHij. This samples
∑j.
I Psips at i die based on .
I Start S = EHF . Once enough psips created change S inresponse to population. e.g. for growth, make S morenegative, so more death.
I Opposite-signed psips at the same determinant annihilate.
ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
Dynamics
∂ci∂τ
= −Hiici −∑j→i
Hijcj.
I At time τ , given a psip at j, at time τ + δτ spawn a new oneat randomly chosen i based on −δτHij. This samples
∑j.
I Psips at i die based on .
I Start S = EHF . Once enough psips created change S inresponse to population. e.g. for growth, make S morenegative, so more death.
I Opposite-signed psips at the same determinant annihilate.
ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
Dynamics
∂ci∂τ
= −Hiici −∑j→i
Hijcj.
I At time τ , given a psip at j, at time τ + δτ spawn a new oneat randomly chosen i based on −δτHij. This samples
∑j.
I Psips at i die based on δτHii.
I Start S = EHF . Once enough psips created change S inresponse to population. e.g. for growth, make S morenegative, so more death.
I Opposite-signed psips at the same determinant annihilate.
ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
Dynamics
∂ci∂τ
= −Hiici −∑j→i
Hijcj. X
I At time τ , given a psip at j, at time τ + δτ spawn a new oneat randomly chosen i based on −δτHij. This samples
∑j.
I Psips at i die based on δτHii.
I Start S = EHF . Once enough psips created change S inresponse to population. e.g. for growth, make S morenegative, so more death.
I Opposite-signed psips at the same determinant annihilate.
ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
Dynamics
∂ci∂τ
= −Hiici −∑j→i
Hijcj.
I At time τ , given a psip at j, at time τ + δτ spawn a new oneat randomly chosen i based on −δτHij. This samples
∑j.
I Psips at i die based on δτHii.
I Start S = EHF . Once enough psips created change S inresponse to population. e.g. for growth, make S morenegative, so more death.
I Opposite-signed psips at the same determinant annihilate.
ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
Dynamics
∂ci∂τ
= − [Hii − S]ci −∑j→i
Hijcj.
I At time τ , given a psip at j, at time τ + δτ spawn a new oneat randomly chosen i based on −δτHij. This samples
∑j.
I Psips at i die based on δτ(Hii − s). Usually S < EHF .
I Start S = EHF . Once enough psips created change S inresponse to population. e.g. for growth, make S morenegative, so more death.
I Opposite-signed psips at the same determinant annihilate.
ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
Dynamics
∂ci∂τ
= − [Hii − S]ci −∑j→i
Hijcj.
I At time τ , given a psip at j, at time τ + δτ spawn a new oneat randomly chosen i based on −δτHij. This samples
∑j.
I Psips at i die based on δτ(Hii − s). Usually S < EHF .
I Start S = EHF . Once enough psips created change S inresponse to population. e.g. for growth, make S morenegative, so more death.
I Opposite-signed psips at the same determinant annihilate.
ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
Dynamics
∂ci∂τ
= − [Hii − S]ci −∑j→i
Hijcj.
I At time τ , given a psip at j, at time τ + δτ spawn a new oneat randomly chosen i based on −δτHij. This samples
∑j.
I Psips at i die based on δτ(Hii − s). Usually S < EHF .
I Start S = EHF . Once enough psips created change S inresponse to population. e.g. for growth, make S morenegative, so more death.
I Opposite-signed psips at the same determinant annihilate.
ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
Dynamics
∂ci∂τ
= − [Hii − S]ci −∑j→i
Hijcj.XXXX
I At time τ , given a psip at j, at time τ + δτ spawn a new oneat randomly chosen i based on −δτHij. This samples
∑j.
I Psips at i die based on δτ(Hii − s). Usually S < EHF .
I Start S = EHF . Once enough psips created change S inresponse to population. e.g. for growth, make S morenegative, so more death.
I Opposite-signed psips at the same determinant annihilate.
ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
Dynamics
∂ci∂τ
= − [Hii − S]ci −∑j→i
Hijcj.
I At time τ , given a psip at j, at time τ + δτ spawn a new oneat randomly chosen i based on −δτHij. This samples
∑j.
I Psips at i die based on δτ(Hii − s). Usually S < EHF .
I Start S = EHF . Once enough psips created change S inresponse to population. e.g. for growth, make S morenegative, so more death.
I Opposite-signed psips at the same determinant annihilate.
ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
Example
I Initialconfiguration
I Spawning
I Death
I Annihilation
I Spawning
I Death
I Annihilation
I Many stepslater
ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
Example
I Initialconfiguration
I Spawning
I Death
I Annihilation
I Spawning
I Death
I Annihilation
I Many stepslater
ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
Example
I Initialconfiguration
I Spawning
I Death
I Annihilation
I Spawning
I Death
I Annihilation
I Many stepslater
X
X
ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
Example
I Initialconfiguration
I Spawning
I Death
I Annihilation
I Spawning
I Death
I Annihilation
I Many stepslater
ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
Example
I Initialconfiguration
I Spawning
I Death
I Annihilation
I Spawning
I Death
I Annihilation
I Many stepslater
ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
Example
I Initialconfiguration
I Spawning
I Death
I Annihilation
I Spawning
I Death
I Annihilation
I Many stepslater
X
X
X
ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
Example
I Initialconfiguration
I Spawning
I Death
I Annihilation
I Spawning
I Death
I Annihilation
I Many stepslater
XX
ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
Example
I Initialconfiguration
I Spawning
I Death
I Annihilation
I Spawning
I Death
I Annihilation
I Many stepslater
ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
Performance
I With a system-dependent critical number of psips, FCIQMCreproduces CI energy essentially exactly.
I Critical psip number, Nc, needed to allow the correct signstructure of the wavefunction to develop.
I Below this, energy is wrong. Plateau shows what Nc is.
0 10000 20000 30000 40000 50000iteration
100
101
102
103
104
105
106
107
# psips
ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
Performance
I With a system-dependent critical number of psips, FCIQMCreproduces CI energy essentially exactly.
I Critical psip number, Nc, needed to allow the correct signstructure of the wavefunction to develop.
I Below this, energy is wrong. Plateau shows what Nc is.
0 10000 20000 30000 40000 50000iteration
100
101
102
103
104
105
106
107
# psips
ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
Performance
I With a system-dependent critical number of psips, FCIQMCreproduces CI energy essentially exactly.
I Critical psip number, Nc, needed to allow the correct signstructure of the wavefunction to develop.
I Below this, energy is wrong. Plateau shows what Nc is.
0 10000 20000 30000 40000 50000iteration
100
101
102
103
104
105
106
107
# psips
ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
Performance
I With a system-dependent critical number of psips, FCIQMCreproduces CI energy essentially exactly.
I Critical psip number, Nc, needed to allow the correct signstructure of the wavefunction to develop.
I Below this, energy is wrong. Plateau shows what Nc is.
0 10000 20000 30000 40000 50000iteration
100
101
102
103
104
105
106
107
# psips
ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
The Initiator Approximation
I Only allow psips on determinants with a significant population(e.g. 3+) to spawn to unoccupied determinants.
I Vastly reduces Nc. Introduces systematic, but controllableerror.
I Scaling exponential with increasing system size — even withinitiators.
I Able to calculate energies for previously insoluble systems(larger by many orders of magnitude).
Deidre Cleland, George H. Booth, and Ali Alavi, J. Chem. Phys. 132, 041103 (2010)
George H. Booth, and Ali Alavi, J. Chem. Phys. 132, 174104 (2010)
Deidre Cleland, George Booth, and Ali Alavi, J. Chem. Phys. 134, 024112 (2011)
ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
The Initiator Approximation
I Only allow psips on determinants with a significant population(e.g. 3+) to spawn to unoccupied determinants.
I Vastly reduces Nc. Introduces systematic, but controllableerror.
I Scaling exponential with increasing system size — even withinitiators.
I Able to calculate energies for previously insoluble systems(larger by many orders of magnitude).
Deidre Cleland, George H. Booth, and Ali Alavi, J. Chem. Phys. 132, 041103 (2010)
George H. Booth, and Ali Alavi, J. Chem. Phys. 132, 174104 (2010)
Deidre Cleland, George Booth, and Ali Alavi, J. Chem. Phys. 134, 024112 (2011)
ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
Excitors
|D0〉
→
aai |D0〉
|D0〉
→
abcjk|D0〉
|D0〉
→
aai abcjk|D0〉 = aabcijk |D0〉
ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
Excitors
|D0〉
→
aai |D0〉 |D0〉
→
abcjk|D0〉
|D0〉
→
aai abcjk|D0〉 = aabcijk |D0〉
ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
Excitors
|D0〉
→
aai |D0〉 |D0〉
→
abcjk|D0〉
|D0〉
→
aai abcjk|D0〉 = aabcijk |D0〉
ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
The exponential Ansatz
I Coupled Cluster Theory uses a cluster operator T =∑
i tiai
I which can be split into excitation levels (and truncated)T = T 1 + T 2 + T 3 + . . .
T 1 =∑ia
tai aai T 2 =
∑i<ja<b
tabij aabij · · ·
I Use an exponential to generate the wavefunction
|ΨCC〉 = eT |D0〉I Exponential generates products of excitors
eT = 1 + T 1 + T 2 + · · ·+ 12 T
21 + 1
2 T 1T 2 + · · ·
ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
The exponential Ansatz
I Coupled Cluster Theory uses a cluster operator T =∑
i tiaiI which can be split into excitation levels (and truncated)T = T 1 + T 2 + T 3 + . . .
T 1 =∑ia
tai aai T 2 =
∑i<ja<b
tabij aabij · · ·
I Use an exponential to generate the wavefunction
|ΨCC〉 = eT |D0〉I Exponential generates products of excitors
eT = 1 + T 1 + T 2 + · · ·+ 12 T
21 + 1
2 T 1T 2 + · · ·
ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
The exponential Ansatz
I Coupled Cluster Theory uses a cluster operator T =∑
i tiaiI which can be split into excitation levels (and truncated)T = T 1 + T 2 + T 3 + . . .
T 1 =∑ia
tai aai T 2 =
∑i<ja<b
tabij aabij · · ·
I Use an exponential to generate the wavefunction
|ΨCC〉 = eT |D0〉I Exponential generates products of excitors
eT = 1 + T 1 + T 2 + · · ·+ 12 T
21 + 1
2 T 1T 2 + · · ·
ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
Size Consistency
I Consider a single atom.
I Excitation aabij is present in both CISD and CCSD.
I Consider two separated non-interacting atoms: X and Y.
I Simultaneous excitation aaXbXaY bYiXjX iY jYneeded for same level of
description of each atom.
I Present in CCSD in 12 T
22 product aaXbXiXjX
aaY bYiY jY.
I Only present in CISDTQ level and beyond.
I As number of electrons increases, truncated CI rapidlyworsens.
ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
Size Consistency
I Consider a single atom.
I Excitation aabij is present in both CISD and CCSD.
I Consider two separated non-interacting atoms: X and Y.
I Simultaneous excitation aaXbXaY bYiXjX iY jYneeded for same level of
description of each atom.
I Present in CCSD in 12 T
22 product aaXbXiXjX
aaY bYiY jY.
I Only present in CISDTQ level and beyond.
I As number of electrons increases, truncated CI rapidlyworsens.
ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
Size Consistency
I Consider a single atom.
I Excitation aabij is present in both CISD and CCSD.
I Consider two separated non-interacting atoms: X and Y.
I Simultaneous excitation aaXbXaY bYiXjX iY jYneeded for same level of
description of each atom.
I Present in CCSD in 12 T
22 product aaXbXiXjX
aaY bYiY jY.
I Only present in CISDTQ level and beyond.
I As number of electrons increases, truncated CI rapidlyworsens.
ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
Size Consistency
I Consider a single atom.
I Excitation aabij is present in both CISD and CCSD.
I Consider two separated non-interacting atoms: X and Y.
I Simultaneous excitation aaXbXaY bYiXjX iY jYneeded for same level of
description of each atom.
I Present in CCSD in 12 T
22 product aaXbXiXjX
aaY bYiY jY.
I Only present in CISDTQ level and beyond.
I As number of electrons increases, truncated CI rapidlyworsens.
ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
Size Consistency
1 2 3 4 5 6Truncation Level
10-3
10-2
10-1
100
101
102
103
Error in Energy per atom / kJ/mol
CC1 kcal/mol
1 kcal/mol = 10 meVˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
Size Consistency
1 2 3 4 5 6Truncation Level
10-3
10-2
10-1
100
101
102
103
Error in Energy per atom / kJ/mol
CC1 kcal/molNe1 CI
1 kcal/mol = 10 meVˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
Size Consistency
1 2 3 4 5 6Truncation Level
10-3
10-2
10-1
100
101
102
103
Error in Energy per atom / kJ/mol
CC1 kcal/molNe1 CI
Ne2 CI
1 kcal/mol = 10 meVˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
Size Consistency
1 2 3 4 5 6Truncation Level
10-3
10-2
10-1
100
101
102
103
Error in Energy per atom / kJ/mol
CC1 kcal/molNe1 CI
Ne2 CI
Ne3 CI
1 kcal/mol = 10 meVˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
Naıve Coupled Cluster Theory
I Solve H|ΨCC〉 = E|ΨCC〉.
I Could solve projected equations like
〈D0|H − E|eTD0〉 = 0
〈Dai |H − E|eTD0〉 = 0
〈Dabij |H − E|eTD0〉 = 0, etc.
I Complicated and non-linear, so solve iteratively.
I Not variational.
I Full CI and Full CC are equivalent. C = eT .
I Normally easier to solve 〈Di|e−T (H − E)eT |D0〉 = 0 (moreon this later).
ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
Naıve Coupled Cluster Theory
I Solve H|ΨCC〉 = E|ΨCC〉.I Could solve projected equations like
〈D0|H − E|eTD0〉 = 0
〈Dai |H − E|eTD0〉 = 0
〈Dabij |H − E|eTD0〉 = 0, etc.
I Complicated and non-linear, so solve iteratively.
I Not variational.
I Full CI and Full CC are equivalent. C = eT .
I Normally easier to solve 〈Di|e−T (H − E)eT |D0〉 = 0 (moreon this later).
ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
Naıve Coupled Cluster Theory
I Solve H|ΨCC〉 = E|ΨCC〉.I Could solve projected equations like
〈D0|H − E|eTD0〉 = 0
〈Dai |H − E|eTD0〉 = 0
〈Dabij |H − E|eTD0〉 = 0, etc.
I Complicated and non-linear, so solve iteratively.
I Not variational.
I Full CI and Full CC are equivalent. C = eT .
I Normally easier to solve 〈Di|e−T (H − E)eT |D0〉 = 0 (moreon this later).
ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
CI vs. CC Theory
〈Di|ΨCC〉 − δτ∑j
(Hij − Eδij)〈Dj|ΨCC〉 = 〈Di|ΨCC〉+ 1
〈Di|(H − E)|ΨCI〉 = 0
〈Di|1− δτ(H − E)|〉 = 〈Di|〉
e.g. Di = Dabij . 〈Di|ΨCC〉 = tabij + tai t
bj − tbi taj = ti+O[T 2].
ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
CI vs. CC Theory
〈Di|ΨCC〉 − δτ∑j
(Hij − Eδij)〈Dj|ΨCC〉 = 〈Di|ΨCC〉+ 1
〈Di|(H − E)|ΨCI〉 = 0
〈Di|1− δτ(H − E)|ΨCI〉 = 〈Di|ΨCI〉
e.g. Di = Dabij . 〈Di|ΨCC〉 = tabij + tai t
bj − tbi taj = ti+O[T 2].
ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
CI vs. CC Theory
〈Di|ΨCC〉 − δτ∑j
(Hij − Eδij)〈Dj|ΨCC〉 = 〈Di|ΨCC〉+ 1
〈Di|(H − E)|ΨCI〉 = 0
〈Di|1− δτ(H − E)|ΨCI〉 = 〈Di|ΨCI〉∑j
〈Di|1− δτ(H − E)|cjDj〉 = ci
e.g. Di = Dabij . 〈Di|ΨCC〉 = tabij + tai t
bj − tbi taj = ti+O[T 2].
ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
CI vs. CC Theory
〈Di|ΨCC〉 − δτ∑j
(Hij − Eδij)〈Dj|ΨCC〉 = 〈Di|ΨCC〉+ 1
〈Di|(H − E)|ΨCI〉 = 0
〈Di|1− δτ(H − E)|ΨCI〉 = 〈Di|ΨCI〉∑j
〈Di|1− δτ(H − E)|cjDj〉 = ci
ci − δτ∑j
(Hij − Eδij)cj = ci
e.g. Di = Dabij . 〈Di|ΨCC〉 = tabij + tai t
bj − tbi taj = ti+O[T 2].
ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
CI vs. CC Theory
〈Di|ΨCC〉 − δτ∑j
(Hij − Eδij)〈Dj|ΨCC〉 = 〈Di|ΨCC〉+ 1
〈Di|(H − E)|ΨCI〉 = 0
〈Di|1− δτ(H − E)|ΨCI〉 = 〈Di|ΨCI〉∑j
〈Di|1− δτ(H − E)|cjDj〉 = ci
ci − δτ∑j
(Hij − Eδij)cj = ci
ci(τ)− δτ∑j
(Hij − Eδij)cj(τ) = ci(τ + δτ)
e.g. Di = Dabij . 〈Di|ΨCC〉 = tabij + tai t
bj − tbi taj = ti+O[T 2].
ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
CI vs. CC Theory
〈Di|ΨCC〉 − δτ∑j
(Hij − Eδij)〈Dj|ΨCC〉 = 〈Di|ΨCC〉+ 1
〈Di|(H − E)|ΨCC〉 = 0
〈Di|1− δτ(H − E)|ΨCI〉 = 〈Di|ΨCI〉∑j
〈Di|1− δτ(H − E)|cjDj〉 = ci
ci − δτ∑j
(Hij − Eδij)cj = ci
ci(τ)− δτ∑j
(Hij − Eδij)cj(τ) = ci(τ + δτ)
e.g. Di = Dabij . 〈Di|ΨCC〉 = tabij + tai t
bj − tbi taj = ti+O[T 2].
ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
CI vs. CC Theory
〈Di|ΨCC〉 − δτ∑j
(Hij − Eδij)〈Dj|ΨCC〉 = 〈Di|ΨCC〉+ 1
〈Di|(H − E)|ΨCC〉 = 0
〈Di|1− δτ(H − E)|ΨCC〉 = 〈Di|ΨCC〉∑j
〈Di|1− δτ(H − E)|cjDj〉 = ci
ci − δτ∑j
(Hij − Eδij)cj = ci
ci(τ)− δτ∑j
(Hij − Eδij)cj(τ) = ci(τ + δτ)
e.g. Di = Dabij . 〈Di|ΨCC〉 = tabij + tai t
bj − tbi taj = ti+O[T 2].
ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
CI vs. CC Theory
〈Di|ΨCC〉 − δτ∑j
(Hij − Eδij)〈Dj|ΨCC〉 = 〈Di|ΨCC〉+ 1
〈Di|(H − E)|ΨCC〉 = 0
〈Di|1− δτ(H − E)|ΨCC〉 = 〈Di|ΨCC〉∑j
〈Di|1− δτ(H − E)|Dj〉〈Dj|ΨCC〉 = 〈Di|ΨCC〉
ci − δτ∑j
(Hij − Eδij)cj = ci
ci(τ)− δτ∑j
(Hij − Eδij)cj(τ) = ci(τ + δτ)
e.g. Di = Dabij . 〈Di|ΨCC〉 = tabij + tai t
bj − tbi taj = ti+O[T 2].
ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
CI vs. CC Theory
〈Di|ΨCC〉 − δτ∑j
(Hij − Eδij)〈Dj|ΨCC〉 = 〈Di|ΨCC〉+ 1
〈Di|(H − E)|ΨCC〉 = 0
〈Di|1− δτ(H − E)|ΨCC〉 = 〈Di|ΨCC〉∑j
〈Di|1− δτ(H − E)|Dj〉〈Dj|ΨCC〉 = 〈Di|ΨCC〉
〈Di|ΨCC〉 − δτ∑j
(Hij − Eδij)〈Dj|ΨCC〉 = 〈Di|ΨCC〉
ci(τ)− δτ∑j
(Hij − Eδij)cj(τ) = ci(τ + δτ)
e.g. Di = Dabij . 〈Di|ΨCC〉 = tabij + tai t
bj − tbi taj = ti+O[T 2].
ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
CI vs. CC Theory
〈Di|ΨCC〉 − δτ∑j
(Hij − Eδij)〈Dj|ΨCC〉 = 〈Di|ΨCC〉+ 1
〈Di|(H − E)|ΨCC〉 = 0
〈Di|1− δτ(H − E)|ΨCC〉 = 〈Di|ΨCC〉∑j
〈Di|1− δτ(H − E)|Dj〉〈Dj|ΨCC〉 = 〈Di|ΨCC〉
〈Di|ΨCC〉 − δτ∑j
(Hij − Eδij)〈Dj|ΨCC〉 = 〈Di|ΨCC〉
ci(τ)− δτ∑j
(Hij − Eδij)cj(τ) = ci(τ + δτ)
e.g. Di = Dabij . 〈Di|ΨCC〉 = tabij + tai t
bj − tbi taj = ti+O[T 2].
ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
CI vs. CC Theory
〈Di|ΨCC〉 − δτ∑j
(Hij − Eδij)〈Dj|ΨCC〉 = 〈Di|ΨCC〉+ 1
〈Di|(H − E)|ΨCC〉 = 0
〈Di|1− δτ(H − E)|ΨCC〉 = 〈Di|ΨCC〉∑j
〈Di|1− δτ(H − E)|Dj〉〈Dj|ΨCC〉 = 〈Di|ΨCC〉
〈Di|ΨCC〉 − δτ∑j
(Hij − Eδij)〈Dj|ΨCC〉 = 〈Di|ΨCC〉
ti(τ)− δτ∑j
(Hij − Eδij)〈Dj|ΨCC(τ)〉 = ti(τ + δτ)
e.g. Di = Dabij . 〈Di|ΨCC〉 = tabij + tai t
bj − tbi taj = ti+O[T 2].
ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
Sampling the CC equations
ti(τ)− δτ∑j
(Hij − Eδij)〈Dj|ΨCC(τ)〉 = ti(τ + δτ)
.I ∀i, ∀j, calculate 〈Dj|ΨCC(τ)〉 and update ti.
I ∀j, calculate 〈Dj|ΨCC(τ)〉. Now randomly pick i, and updateti.
eT = 1 +∑i
tiai +1
2!
∑i,j
titjaiaj +1
3!
∑i,j,k
titjtkaiajak + . . .
.I Pick a random term in eT and work out j and 〈Dj|ΨCC(τ)〉
from just this term. Now randomly pick i, and update ti.
ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
Sampling the CC equations
ti(τ)− δτ∑j
(Hij − Eδij)〈Dj|ΨCC(τ)〉 = ti(τ + δτ)
.I ∀i, ∀j, calculate 〈Dj|ΨCC(τ)〉 and update ti.I ∀j, calculate 〈Dj|ΨCC(τ)〉. Now randomly pick i, and updateti.
eT = 1 +∑i
tiai +1
2!
∑i,j
titjaiaj +1
3!
∑i,j,k
titjtkaiajak + . . .
.I Pick a random term in eT and work out j and 〈Dj|ΨCC(τ)〉
from just this term. Now randomly pick i, and update ti.
ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
Sampling the CC equations
ti(τ)− δτ∑j
(Hij − Eδij)〈Dj|ΨCC(τ)〉 = ti(τ + δτ)
.I ∀i, ∀j, calculate 〈Dj|ΨCC(τ)〉 and update ti.I ∀j, calculate 〈Dj|ΨCC(τ)〉. Now randomly pick i, and updateti.
eT = 1 +∑i
tiai +1
2!
∑i,j
titjaiaj +1
3!
∑i,j,k
titjtkaiajak + . . .
.
I Pick a random term in eT and work out j and 〈Dj|ΨCC(τ)〉from just this term. Now randomly pick i, and update ti.
ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
Sampling the CC equations
ti(τ)− δτ∑j
(Hij − Eδij)〈Dj|ΨCC(τ)〉 = ti(τ + δτ)
.I ∀i, ∀j, calculate 〈Dj|ΨCC(τ)〉 and update ti.I ∀j, calculate 〈Dj|ΨCC(τ)〉. Now randomly pick i, and updateti.
eT = 1 +∑i
tiai +1
2!
∑i,j
titjaiaj +1
3!
∑i,j,k
titjtkaiajak + . . .
.I Pick a random term in eT and work out j and 〈Dj|ΨCC(τ)〉
from just this term. Now randomly pick i, and update ti.
ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
Discretization
I Represent T =∑
i tiai as a population of excips for eachexcitation.
I Need to sample
eT = 1 +∑i
tiai +1
2!
∑i,j
titjaiaj +1
3!
∑i,j,k
titjtkaiajak + . . .
I From list of excips randomly select a number of them(remembering signs). Need normalized probabilities.
+ai +b
i +bi −aj −bj +c
j +cj +c
j −abij +acik −abcijk
AJWT Phys. Rev. Lett. 105, 263004 (2010)
ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
Discretization
I Represent T =∑
i tiai as a population of excips for eachexcitation.
I Need to sample
eT = 1 +∑i
tiai +1
2!
∑i,j
titjaiaj +1
3!
∑i,j,k
titjtkaiajak + . . .
I From list of excips randomly select a number of them(remembering signs). Need normalized probabilities.
+ai +b
i +bi −aj −bj +c
j +cj +c
j −abij +acik −abcijk
AJWT Phys. Rev. Lett. 105, 263004 (2010)
ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
Discretization
I Represent T =∑
i tiai as a population of excips for eachexcitation.
I Need to sample
eT = 1 +∑i
tiai +1
2!
∑i,j
titjaiaj +1
3!
∑i,j,k
titjtkaiajak + . . .
I From list of excips randomly select a number of them(remembering signs). Need normalized probabilities.
+ai +b
i +bi −aj −bj +c
j +cj +c
j −abij +acik −abcijk
AJWT Phys. Rev. Lett. 105, 263004 (2010)
ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
Discretization
I Represent T =∑
i tiai as a population of excips for eachexcitation.
I Need to sample
eT = 1 +∑i
tiai +1
2!
∑i,j
titjaiaj +1
3!
∑i,j,k
titjtkaiajak + . . .
I From list of excips randomly select a number of them(remembering signs). Need normalized probabilities.
+ai +b
i +bi −aj −bj +c
j +cj +c
j −abij +acik −abcijk
AJWT Phys. Rev. Lett. 105, 263004 (2010)
ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
Discretization
+acik −bj
→ −aabcijk
I Collapse this into a single excitor and apply to the reference,resulting in a determinant.
−aabcijk |D0〉 = −|Dabcijk 〉
I Spawning: Pick random excitation from |Dabcijk 〉 (e.g. c
i ) and
spawn according to −δτ〈Dci |H|Dabc
ijk 〉. . . . Create new +ci
excip.I Death: Die according to δτ(〈Dabc
ijk |H|Dabcijk 〉 − S). . . . Create
new +abcijk excip.
I Annihilate as in FCIQMC.
ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
Discretization
+acik −bj → −aabcijk
I Collapse this into a single excitor and apply to the reference,resulting in a determinant.
−aabcijk |D0〉 = −|Dabcijk 〉
I Spawning: Pick random excitation from |Dabcijk 〉 (e.g. c
i ) and
spawn according to −δτ〈Dci |H|Dabc
ijk 〉. . . . Create new +ci
excip.I Death: Die according to δτ(〈Dabc
ijk |H|Dabcijk 〉 − S). . . . Create
new +abcijk excip.
I Annihilate as in FCIQMC.
ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
Discretization
+acik −bj → −aabcijk
I Collapse this into a single excitor and apply to the reference,resulting in a determinant.
−aabcijk |D0〉 = −|Dabcijk 〉
I Spawning: Pick random excitation from |Dabcijk 〉 (e.g. c
i ) and
spawn according to −δτ〈Dci |H|Dabc
ijk 〉.
. . . Create new +ci
excip.I Death: Die according to δτ(〈Dabc
ijk |H|Dabcijk 〉 − S). . . . Create
new +abcijk excip.
I Annihilate as in FCIQMC.
ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
Discretization
+acik −bj → −aabcijk
I Collapse this into a single excitor and apply to the reference,resulting in a determinant.
−aabcijk |D0〉 = −|Dabcijk 〉
I Spawning: Pick random excitation from |Dabcijk 〉 (e.g. c
i ) and
spawn according to −δτ〈Dci |H|Dabc
ijk 〉. . . . Create new +ci
excip.
I Death: Die according to δτ(〈Dabcijk |H|Dabc
ijk 〉 − S). . . . Create
new +abcijk excip.
I Annihilate as in FCIQMC.
ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
Discretization
+acik −bj → −aabcijk
I Collapse this into a single excitor and apply to the reference,resulting in a determinant.
−aabcijk |D0〉 = −|Dabcijk 〉
I Spawning: Pick random excitation from |Dabcijk 〉 (e.g. c
i ) and
spawn according to −δτ〈Dci |H|Dabc
ijk 〉. . . . Create new +ci
excip.I Death: Die according to δτ(〈Dabc
ijk |H|Dabcijk 〉 − S).
. . . Create
new +abcijk excip.
I Annihilate as in FCIQMC.
ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
Discretization
+acik −bj → −aabcijk
I Collapse this into a single excitor and apply to the reference,resulting in a determinant.
−aabcijk |D0〉 = −|Dabcijk 〉
I Spawning: Pick random excitation from |Dabcijk 〉 (e.g. c
i ) and
spawn according to −δτ〈Dci |H|Dabc
ijk 〉. . . . Create new +ci
excip.I Death: Die according to δτ(〈Dabc
ijk |H|Dabcijk 〉 − S). . . . Create
new +abcijk excip.
I Annihilate as in FCIQMC.
ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
Discretization
+acik −bj → −aabcijk
I Collapse this into a single excitor and apply to the reference,resulting in a determinant.
−aabcijk |D0〉 = −|Dabcijk 〉
I Spawning: Pick random excitation from |Dabcijk 〉 (e.g. c
i ) and
spawn according to −δτ〈Dci |H|Dabc
ijk 〉. . . . Create new +ci
excip.I Death: Die according to δτ(〈Dabc
ijk |H|Dabcijk 〉 − S). . . . Create
new +abcijk excip.
I Annihilate as in FCIQMC.ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
Coupled Cluster Monte Carlo
ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
Performance
I Initial excip growth much faster than FCIQMC.
I Critical number of excips required otherwise growth unstable.
I Critical number of excips seems to scale with size of space.e.g. CCSD Nc ∼ N4.
I Able to reproduce large coupled cluster calculations in lesstime.
I Initiator approximation also applicable.
I Cluster is initiator iff all component excips have sufficientamplitude.
I Much fewer excips, so much faster.
ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
Performance
I Initial excip growth much faster than FCIQMC.
I Critical number of excips required otherwise growth unstable.
I Critical number of excips seems to scale with size of space.e.g. CCSD Nc ∼ N4.
I Able to reproduce large coupled cluster calculations in lesstime.
I Initiator approximation also applicable.
I Cluster is initiator iff all component excips have sufficientamplitude.
I Much fewer excips, so much faster.
ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
Performance
I Initial excip growth much faster than FCIQMC.
I Critical number of excips required otherwise growth unstable.
I Critical number of excips seems to scale with size of space.e.g. CCSD Nc ∼ N4.
I Able to reproduce large coupled cluster calculations in lesstime.
I Initiator approximation also applicable.
I Cluster is initiator iff all component excips have sufficientamplitude.
I Much fewer excips, so much faster.
ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
H2O cc-pVDZ
−0.24
−0.22
−0.20
−0.18
−0.16
−0.14
−0.12
Ene
rgy
/Har
trees ECCSDTQ
〈E(τ)〉〈E(τ)〉ma
0 1000 2000 3000 4000 5000Iterations
10−6
10−5
10−4
10−3
10−2
10−1
Ene
rgy
Err
or/H
artre
es
|〈E(τ)〉ma − ECCSDTQ||〈E(τ)〉equil − ECCSDTQ|
ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
Ne cc-pVQZ initiator
0 20000 40000 60000 80000 100000 120000 140000 160000# excips
−0.3365
−0.3360
−0.3355
−0.3350
−0.3345
−0.3340
−0.3335
−0.3330C
orr
ela
tion E
nerg
y/H
ar
ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
Scaling
TZ QZ 5Z 6ZBasis
104
105
106
107
108
109
CPU Time / s
minute
day
week
year
decade
millennium
239Pu half life
aeon
age of universe
Full CCSDTQ timeMC CCSDTQ time 105
106
107
108
109
1010
1011
Storage / bytes
Full CCSDTQ memMC CCSDTQ mem
ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
Parallel Scaling
10 20 30 40 50 60 70 80 90 100Cores
0.5
1.0
1.5
2.0
2.5
3.0
3.5
4.0
4.5
Speedup rel. to 24
Ideal ScalingCCMC
ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
The Uniform Electron Gas
rs = 1
186−1342−1514−11030−14218−1
1/# basis functions
−1.10
−1.05
−1.00
−0.95
−0.90
Corr
ela
tion E
nerg
y/e
lect
ron /
eV n14 iFCIQMC
n = 14 Cutoff 64Ry, M = 4218, triples: 3× 108, full 1034
n = 54 Cutoff 64Ry, M = 4218, triples: 2× 1010, full 10117J. J. Shepherd, G. H. Booth and A. Alavi J. Chem. Phys. 136, 244101 (2012)
J. J. Shepherd, A. Gruneis, G. H. Booth, G. Kresse, and A. Alavi Phys. Rev. B 86, 035111 (2012)
ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
The Uniform Electron Gas
rs = 1
186−1342−1514−11030−14218−1
1/# basis functions
−1.10
−1.05
−1.00
−0.95
−0.90
Corr
ela
tion E
nerg
y/e
lect
ron /
eV n14 iFCIQMC
n14 CCSD
n = 14 Cutoff 64Ry, M = 4218, triples: 3× 108, full 1034
n = 54 Cutoff 64Ry, M = 4218, triples: 2× 1010, full 10117J. J. Shepherd, G. H. Booth and A. Alavi J. Chem. Phys. 136, 244101 (2012)
J. J. Shepherd, A. Gruneis, G. H. Booth, G. Kresse, and A. Alavi Phys. Rev. B 86, 035111 (2012)
ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
The Uniform Electron Gas
rs = 1
186−1342−1514−11030−14218−1
1/# basis functions
−1.10
−1.05
−1.00
−0.95
−0.90
Corr
ela
tion E
nerg
y/e
lect
ron /
eV n14 iFCIQMC
n14 CCSDn14 iCCSDT
n = 14 Cutoff 64Ry, M = 4218, triples: 3× 108, full 1034
n = 54 Cutoff 64Ry, M = 4218, triples: 2× 1010, full 10117J. J. Shepherd, G. H. Booth and A. Alavi J. Chem. Phys. 136, 244101 (2012)
J. J. Shepherd, A. Gruneis, G. H. Booth, G. Kresse, and A. Alavi Phys. Rev. B 86, 035111 (2012)
ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
The Uniform Electron Gas
rs = 1
186−1342−1514−11030−14218−1
1/# basis functions
−1.1
−1.0
−0.9
−0.8
−0.7
−0.6C
orr
ela
tion E
nerg
y/e
lect
ron /
eV DMC
n54 iFCIQMC
n = 14 Cutoff 64Ry, M = 4218, triples: 3× 108, full 1034
n = 54 Cutoff 64Ry, M = 4218, triples: 2× 1010, full 10117J. J. Shepherd, G. H. Booth and A. Alavi J. Chem. Phys. 136, 244101 (2012)
J. J. Shepherd, A. Gruneis, G. H. Booth, G. Kresse, and A. Alavi Phys. Rev. B 86, 035111 (2012)
ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
The Uniform Electron Gas
rs = 1
186−1342−1514−11030−14218−1
1/# basis functions
−1.1
−1.0
−0.9
−0.8
−0.7
−0.6C
orr
ela
tion E
nerg
y/e
lect
ron /
eV DMC
n54 iFCIQMCn54 CCSD MC
n = 14 Cutoff 64Ry, M = 4218, triples: 3× 108, full 1034
n = 54 Cutoff 64Ry, M = 4218, triples: 2× 1010, full 10117J. J. Shepherd, G. H. Booth and A. Alavi J. Chem. Phys. 136, 244101 (2012)
J. J. Shepherd, A. Gruneis, G. H. Booth, G. Kresse, and A. Alavi Phys. Rev. B 86, 035111 (2012)
ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
The Uniform Electron Gas
rs = 1
186−1342−1514−11030−14218−1
1/# basis functions
−1.1
−1.0
−0.9
−0.8
−0.7
−0.6C
orr
ela
tion E
nerg
y/e
lect
ron /
eV DMC
n54 iFCIQMCn54 CCSD MCn54 iCCSD MC
n = 14 Cutoff 64Ry, M = 4218, triples: 3× 108, full 1034
n = 54 Cutoff 64Ry, M = 4218, triples: 2× 1010, full 10117J. J. Shepherd, G. H. Booth and A. Alavi J. Chem. Phys. 136, 244101 (2012)
J. J. Shepherd, A. Gruneis, G. H. Booth, G. Kresse, and A. Alavi Phys. Rev. B 86, 035111 (2012)
ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
The Uniform Electron Gas
10-1 100 101
rs / a0
−1.4
−1.2
−1.0
−0.8
−0.6
−0.4
−0.2C
orr
ela
tion E
nerg
y/e
lect
ron /
eV n14 FCI
J. J. Shepherd, G. H. Booth and A. Alavi J. Chem. Phys. 136, 244101 (2012)
J. J. Shepherd, A. Gruneis, G. H. Booth, G. Kresse, and A. Alavi Phys. Rev. B 86, 035111 (2012)
ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
The Uniform Electron Gas
10-1 100 101
rs / a0
−1.4
−1.2
−1.0
−0.8
−0.6
−0.4
−0.2C
orr
ela
tion E
nerg
y/e
lect
ron /
eV n14 FCI
n14 MP2
J. J. Shepherd, G. H. Booth and A. Alavi J. Chem. Phys. 136, 244101 (2012)J. J. Shepherd, A. Gruneis, G. H. Booth, G. Kresse, and A. Alavi Phys. Rev. B 86, 035111 (2012)
ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
The Uniform Electron Gas
10-1 100 101
rs / a0
−1.4
−1.2
−1.0
−0.8
−0.6
−0.4
−0.2C
orr
ela
tion E
nerg
y/e
lect
ron /
eV n14 FCI
n14 MP2n14 CCSD
J. J. Shepherd, G. H. Booth and A. Alavi J. Chem. Phys. 136, 244101 (2012)J. J. Shepherd, A. Gruneis, G. H. Booth, G. Kresse, and A. Alavi Phys. Rev. B 86, 035111 (2012)
ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
The Uniform Electron Gas
10-1 100 101
rs / a0
−1.4
−1.2
−1.0
−0.8
−0.6
−0.4
−0.2C
orr
ela
tion E
nerg
y/e
lect
ron /
eV n14 FCI
n14 MP2n14 CCSDn14 CCSDT
J. J. Shepherd, G. H. Booth and A. Alavi J. Chem. Phys. 136, 244101 (2012)J. J. Shepherd, A. Gruneis, G. H. Booth, G. Kresse, and A. Alavi Phys. Rev. B 86, 035111 (2012)
ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
The Uniform Electron Gas
10-1 100 101
rs / a0
−1.4
−1.2
−1.0
−0.8
−0.6
−0.4
−0.2C
orr
ela
tion E
nerg
y/e
lect
ron /
eV n14 RPA+SOSEX
n14 FCIn14 MP2n14 CCSDn14 CCSDT
J. J. Shepherd, G. H. Booth and A. Alavi J. Chem. Phys. 136, 244101 (2012)J. J. Shepherd, A. Gruneis, G. H. Booth, G. Kresse, and A. Alavi Phys. Rev. B 86, 035111 (2012)
ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
The Uniform Electron Gas
10-1 100 101
rs / a0
−1.4
−1.2
−1.0
−0.8
−0.6
−0.4
−0.2C
orr
ela
tion E
nerg
y/e
lect
ron /
eV n54 iCCSD
P. Lopez Rıos, A. Ma, N. D. Drummond, M. D. Towler, and R. J. Needs Phys. Rev. E 74, 066701 (2006)J. J. Shepherd, A. Gruneis, G. H. Booth and A. Alavi Phys. Rev. B 85, 081103(R) (2012)
ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
The Uniform Electron Gas
10-1 100 101
rs / a0
−1.4
−1.2
−1.0
−0.8
−0.6
−0.4
−0.2C
orr
ela
tion E
nerg
y/e
lect
ron /
eV n54 iCCSD
n54 BF DMC
P. Lopez Rıos, A. Ma, N. D. Drummond, M. D. Towler, and R. J. Needs Phys. Rev. E 74, 066701 (2006)
J. J. Shepherd, A. Gruneis, G. H. Booth and A. Alavi Phys. Rev. B 85, 081103(R) (2012)
ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
The Uniform Electron Gas
10-1 100 101
rs / a0
−1.4
−1.2
−1.0
−0.8
−0.6
−0.4
−0.2C
orr
ela
tion E
nerg
y/e
lect
ron /
eV n54 iCCSD
n54 BF DMCn54 FCIQMC
P. Lopez Rıos, A. Ma, N. D. Drummond, M. D. Towler, and R. J. Needs Phys. Rev. E 74, 066701 (2006)J. J. Shepherd, A. Gruneis, G. H. Booth and A. Alavi Phys. Rev. B 85, 081103(R) (2012)
ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
Linked Coupled Cluster Thoery
〈Di|(H − E)eT |D0〉 = 0. Unlinked
〈Di|e−T (H − E)eT |D0〉 = 0. Linked
ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
Linked Coupled Cluster Thoery
〈Di|(H − E)eT |D0〉 = 0. Unlinked
〈Di|e−T (H − E)eT |D0〉 = 0. LinkedBaker–Campbell–Hausdorff expansion simplifies this significantly.
e−T HeT = H + [H, T ] +1
2![[H, T ], T ] +
1
3![[[H, T ], T ], T ]
+1
4![[[[H, T ], T ], T ], T ] + · · ·
ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
Linked Coupled Cluster Thoery
〈Di|(H − E)eT |D0〉 = 0. Unlinked
〈Di|e−T (H − E)eT |D0〉 = 0. LinkedBaker–Campbell–Hausdorff expansion simplifies this significantly.
e−T HeT = H + [H, T ] +1
2![[H, T ], T ] +
1
3![[[H, T ], T ], T ]
+1
4![[[[H, T ], T ], T ], T ]
ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
Linked Coupled Cluster Thoery
〈Di|(H − E)eT |D0〉 = 0. Unlinked
〈Di|e−T (H − E)eT |D0〉 = 0. LinkedBaker–Campbell–Hausdorff expansion simplifies this significantly.
e−T HeT = H + [H, T ] +1
2![[H, T ], T ] +
1
3![[[H, T ], T ], T ]
+1
4![[[[H, T ], T ], T ], T ]
ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
Linked Coupled Cluster Thoery
〈Di|(H − E)eT |D0〉 = 0. Unlinked
〈Di|e−T (H − E)eT |D0〉 = 0. LinkedBaker–Campbell–Hausdorff expansion simplifies this significantly.
e−T HeT = H + [H, T ] +1
2![[H, T ], T ] +
1
3![[[H, T ], T ], T ]
+1
4![[[[H, T ], T ], T ], T ]
[[H, T ], T ] = HT T − T HT − T HT + T T H
ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
Linked Coupled Cluster Thoery
〈Di|(H − E)eT |D0〉 = 0. Unlinked
〈Di|e−T (H − E)eT |D0〉 = 0. LinkedBaker–Campbell–Hausdorff expansion simplifies this significantly.
e−T HeT = H + [H, T ] +1
2![[H, T ], T ] +
1
3![[[H, T ], T ], T ]
+1
4![[[[H, T ], T ], T ], T ]
[[H, T ], T ] = HT T − T HT − T HT + T T H[[H, tαaα], tβ aβ] = tαtβ(Haαaβ − aαHaβ − aβHaα + aαaβH
ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
Linked Coupled Cluster Thoery
〈Di|(H − E)eT |D0〉 = 0. Unlinked
〈Di|e−T (H − E)eT |D0〉 = 0. LinkedBaker–Campbell–Hausdorff expansion simplifies this significantly.
e−T HeT = H + [H, T ] +1
2![[H, T ], T ] +
1
3![[[H, T ], T ], T ]
+1
4![[[[H, T ], T ], T ], T ]
[[H, T ], T ] = HT T − T HT − T HT + T T H[[H, tαaα], tβ aβ] = tαtβ(Haαaβ − aαHaβ − aβHaα + aαaβH
〈Di|
tαtβ(hγ aγ aαaβ − aαHaβ − aβHaα + aαaβH|D0〉Pick a single spawning, sampling H.
ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
Linked Coupled Cluster Thoery
〈Di|(H − E)eT |D0〉 = 0. Unlinked
〈Di|e−T (H − E)eT |D0〉 = 0. LinkedBaker–Campbell–Hausdorff expansion simplifies this significantly.
e−T HeT = H + [H, T ] +1
2![[H, T ], T ] +
1
3![[[H, T ], T ], T ]
+1
4![[[[H, T ], T ], T ], T ]
[[H, T ], T ] = HT T − T HT − T HT + T T H[[H, tαaα], tβ aβ] = tαtβ(Haαaβ − aαHaβ − aβHaα + aαaβH
〈Di|tαtβ(hγ aγ aαaβ − aαHaβ − aβHaα + aαaβH|D0〉Pick a single spawning, sampling H. Determines Di.
ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
Linked Coupled Cluster Thoery
〈Di|(H − E)eT |D0〉 = 0. Unlinked
〈Di|e−T (H − E)eT |D0〉 = 0. LinkedBaker–Campbell–Hausdorff expansion simplifies this significantly.
e−T HeT = H + [H, T ] +1
2![[H, T ], T ] +
1
3![[[H, T ], T ], T ]
+1
4![[[[H, T ], T ], T ], T ]
[[H, T ], T ] = HT T − T HT − T HT + T T H[[H, tαaα], tβ aβ] = tαtβ(Haαaβ − aαHaβ − aβHaα + aαaβH
〈Di|tαtβ(hγ aγ aαaβ − aαhγ aγ aβ − aβhγ aγ aα + aαaβhγ aγ |D0〉Pick a single spawning, sampling H. Determines Di. Use the sameexcitation for H in all terms.
ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
Linked Coupled Cluster Thoery
Pros:
I Exact cancellation should reduce plateaux and thus difficulty.
I Energy (=Shift) is decoupled from excitor population.
I Only need clusters of up to four excitors — should scale muchbetter
Cons:
I Far less simple: Lots of terms to code up.
ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
Conclusions
I Stochastic Coupled Cluster (CCMC) behaves very similar toFCIQMC
I Truncated CC allows polynomial scaling spaces to be used.
I Much simpler to implement than deterministic CC.
I Arbitrary truncation (e.g. CCSDTQ5) uses identical code.
I Required number of excips appears to scale with size of space.
I Initiator approximation very successful at reducing criticalnumber of excips.
I Feasible on workstations, and very parallelizable.
ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
Directions
I Solids. Complex excips. (Volunteer applications welcomed!)
I Large-scale parallelization.
I f12-CCMC.
I CASCC.
I Excitation Energies.
I Available in MOLPRO (Alavi group’s NECI code).
I New open-source code, HANDE, now available from ImperialCollege.
I Applications...
ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
Directions
I Solids. Complex excips. (Volunteer applications welcomed!)
I Large-scale parallelization.
I f12-CCMC.
I CASCC.
I Excitation Energies.
I Available in MOLPRO (Alavi group’s NECI code).
I New open-source code, HANDE, now available from ImperialCollege.
I Applications...
ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory
Acknowledgements
MAGDALENE COLLEGECAMBRIDGE
I Ali AlaviI George Booth — FCIQMCI Deirdre Cleland — InitiatorsI James Shepherd — FCIQMC UEGI James Spencer — Open Source implementationI Alberto Zoccante — Linked CC
ˆ FCIQMC Coupled Cluster Coupled Cluster Monte Carlo $
Alex ThomStochastic Coupled Cluster Theory