Symbolic and numeric Scientific computing for atomic …msafrono/Safronova2007.pdf · Symbolic and...
Transcript of Symbolic and numeric Scientific computing for atomic …msafrono/Safronova2007.pdf · Symbolic and...
Symbolic and numeric Symbolic and numeric Symbolic and numeric Symbolic and numeric
Scientific computing for Scientific computing for Scientific computing for Scientific computing for
atomic phySicSatomic phySicSatomic phySicSatomic phySicS
Symbolic and numeric Symbolic and numeric Symbolic and numeric Symbolic and numeric
Scientific computing for Scientific computing for Scientific computing for Scientific computing for
atomic phySicSatomic phySicSatomic phySicSatomic phySicS
University of Delaware University of Delaware Computational Science Initiative meetingComputational Science Initiative meeting
November 9, 2007
Marianna SafronovaMarianna Safronova
• Motivation
• Symbolic computing
• Numeric computing
• Computational challenges & future projects
outlineoutlineoutlineoutlineoutlineoutlineoutlineoutline
Research field:Research field: development of development of highhigh --precision methodologies for precision methodologies for the calculation of the the calculation of the atomic atomic propertiesproperties and applications of and applications of such calculations. such calculations.
reSearch fieldreSearch fieldreSearch fieldreSearch fieldreSearch fieldreSearch fieldreSearch fieldreSearch field
This research involves:This research involves:
•• the study of thethe study of the fundamental physics fundamental physics problems problems
•• applications of atomic physics to future applications of atomic physics to future technological technological developmentsdevelopments
http://CPEPweb.org
motivation: fundamental phySicSmotivation: fundamental phySicSmotivation: fundamental phySicSmotivation: fundamental phySicSmotivation: fundamental phySicSmotivation: fundamental phySicSmotivation: fundamental phySicSmotivation: fundamental phySicS
http://CPEPweb.org, http://public.web.cern.ch/, Cs experiment, University of Colorado
http://CPEPweb.org
motivation: motivation: motivation: motivation: applicationS to applicationS to applicationS to applicationS to
variouS fieldS of phySicSvariouS fieldS of phySicSvariouS fieldS of phySicSvariouS fieldS of phySicS
motivation: motivation: motivation: motivation: applicationS to applicationS to applicationS to applicationS to
variouS fieldS of phySicSvariouS fieldS of phySicSvariouS fieldS of phySicSvariouS fieldS of phySicS
• Search for variation of the fundamental fine-structure constant with time
• Development of the high-precision methodologies , benchmark problems for understanding of atomic properties, tests of ab initio theory and experimental methods, and cross-checks of experiment
• Determination of nuclear properties such as nuclear magnetic and anapole moments
• Providing data for astrophysics studies
Quantum communication, cryptography Quantum communication, cryptography Quantum communication, cryptography Quantum communication, cryptography
and Quantum information proceSSingand Quantum information proceSSingand Quantum information proceSSingand Quantum information proceSSing
motivation: motivation: motivation: motivation:
new technologieSnew technologieSnew technologieSnew technologieS
motivation: motivation: motivation: motivation:
new technologieSnew technologieSnew technologieSnew technologieS
S1/2
P1/2
D5/2
Quantumbit
Quantum Computer(Innsbruck)
motiation: next motiation: next motiation: next motiation: next
generation atomic clockSgeneration atomic clockSgeneration atomic clockSgeneration atomic clockS
motiation: next motiation: next motiation: next motiation: next
generation atomic clockSgeneration atomic clockSgeneration atomic clockSgeneration atomic clockS
Next - generation ultra precise atomic clock
Atoms trapped by laser light
http://CPEPweb.org
The ability to develop more precise optical frequency standards will open ways to improve global positioningsystem (GPS) measurements and tracking of deep-spaceprobes, perform more accurate measurements of the physical constants and tests of fundamental physics such as searches for gravitational waves, etc.
motivationmotivationmotivationmotivationmotivationmotivationmotivationmotivation
AtomicClocks
S1/2
P1/2D5/2
„quantumbit“
Parity Violation
Quantum information
NEEDNEEDATOMIC ATOMIC PROPERTIESPROPERTIES
calculation of atomic calculation of atomic calculation of atomic calculation of atomic
properiteSproperiteSproperiteSproperiteS
calculation of atomic calculation of atomic calculation of atomic calculation of atomic
properiteSproperiteSproperiteSproperiteS
Step 1 Derivation of the formulas
Step 2 Conducting possible analytical calculationsto simplify expressions for numeric evaluation
Step 3 Writing and debugging the programs for numeric calculations
Step 4 Numeric calculations
calculation of atomic calculation of atomic calculation of atomic calculation of atomic
properiteSproperiteSproperiteSproperiteS
calculation of atomic calculation of atomic calculation of atomic calculation of atomic
properiteSproperiteSproperiteSproperiteS
Step 1 Derivation of the formulas
Step 2 Conducting possible analytical calculationsto simplify expression for numerical evaluation
Step 3 Writing and debugging the programs for numerical calculations
Step 4 Numeric computation
Steps 1 Steps 1 –– 3: Symbolic computing
3: Symbolic computing
coupledcoupledcoupledcoupled----cluSter methodcluSter methodcluSter methodcluSter methodcoupledcoupledcoupledcoupled----cluSter methodcluSter methodcluSter methodcluSter method
The coupledcoupledcoupledcoupled----cluster methodcluster methodcluster methodcluster method sums infinite setsinfinite setsinfinite setsinfinite sets of many-body perturbation theory terms. The wave function of the valence electron v is represented as an expansion that includes all possible single, double, and partial triple excitations.
1s22s22p63s23p63d104s24p64d105s25p66s
1s1s22…5p…5p66 6s6scorevalenceelectron
Cs: atom with single (valence) electron outside of a closed core.
coupledcoupledcoupledcoupled----cluSter methodcluSter methodcluSter methodcluSter methodcoupledcoupledcoupledcoupled----cluSter methodcluSter methodcluSter methodcluSter method
vΨ = Lowest orderCore
core
valence electron
any excited orbital
Single-particle excitations
Double-particle excitations
(0)vΨ
† (0)m m
ma a v
aa aρ+ Ψ∑
† (0)v v vm m
m va aρ
≠+ Ψ∑
† † (0) (0† † )12
...
vm n mna a aa m nmnmnmn
b b v vaab
va a aa a aa aρ ρ+ Ψ + Ψ
+
∑ ∑
2 (0)2 | | 0: :
12
: cm n c adi j l k bv vr sa a aa a aa a a aaH aS a+ + + ++ ++ >→Ψ >
800 terms!800 terms!800 terms!800 terms!800 terms!800 terms!800 terms!800 terms!
Step 1: derivation of the formulaSStep 1: derivation of the formulaSStep 1: derivation of the formulaSStep 1: derivation of the formulaSStep 1: derivation of the formulaSStep 1: derivation of the formulaSStep 1: derivation of the formulaSStep 1: derivation of the formulaS
Need for symbolic computingNeed for symbolic computing
The code was developed to implement Wick’s theorem and simplify the expression.
Symbolic program for coupledSymbolic program for coupledSymbolic program for coupledSymbolic program for coupled----cluSter cluSter cluSter cluSter
method & perturbation theorymethod & perturbation theorymethod & perturbation theorymethod & perturbation theory
Symbolic program for coupledSymbolic program for coupledSymbolic program for coupledSymbolic program for coupled----cluSter cluSter cluSter cluSter
method & perturbation theorymethod & perturbation theorymethod & perturbation theorymethod & perturbation theory
Input: expression of the type
in ASCII format.
Output: simplified resulting formula in the LaTex format, ASCII output is also generated.
:: : :ijkl mnwa i j l m wak ng a a aaa aaaρ + ++ + +
program featureSprogram featureSprogram featureSprogram featureSprogram featureSprogram featureSprogram featureSprogram featureS
1) The code is set to work with two or three normal products (all possible cases) with large number of operators.
2) The code differentiates between different types of indices, i.e.core (a,b,c,…), valence (v,w,x,y,...), excited orbitals (m,n,r,s,…), and general case (i, j, k, l,…).
3) The operators are ordered as required in the same order for all terms.
4) The expression is simplified to account for the identical terms and symmetry rules, .
5) The direct and exchange terms are joined together, .
ijkl jilkgg =
ijkl ijkl ijlkg gg = −%
mnba rswc m n r s a c wbg a a a a a a a aρ + + + + +
Step 2: analytical calculationSStep 2: analytical calculationSStep 2: analytical calculationSStep 2: analytical calculationSStep 2: analytical calculationSStep 2: analytical calculationSStep 2: analytical calculationSStep 2: analytical calculationS
Certain summations can be carried out analytically
Two symbolic programs are currently available:
1) KENTARO – currently interfaced with the previousprogram.
2) RACAH Maple code – very convenient to use, but hascompatibility problems.
Step 3: Symbolic computing for Step 3: Symbolic computing for Step 3: Symbolic computing for Step 3: Symbolic computing for
program generationprogram generationprogram generationprogram generation
Step 3: Symbolic computing for Step 3: Symbolic computing for Step 3: Symbolic computing for Step 3: Symbolic computing for
program generationprogram generationprogram generationprogram generation
The resulting expressions that need to be evaluated numerically contain very large number of terms, resulting in tedious coding and debugging.
The symbolic program generator was developed for this purpose to automatically generate efficient numerical codes for coupled-cluster or perturbation theory terms.
( )
1 2 3 '
2 ' ' 3 11 2 3
1 2 3
' '
( 1)
( ) ( ' ' ) ( )
c d a b c dJ k k k j j j jc d w w
abcd k k k w v w v b a
k k k
c d v w a b
J j j J j j J j j
k j j k j j k j j
X cdab X v w cd Z vwab
ε ε ε ε ε ε ε
+ + + + + + + − −
× + − − + −
∑ ∑
( )' 'v wε−
Step 3: Symbolic computing for Step 3: Symbolic computing for Step 3: Symbolic computing for Step 3: Symbolic computing for
program generationprogram generationprogram generationprogram generation
Step 3: Symbolic computing for Step 3: Symbolic computing for Step 3: Symbolic computing for Step 3: Symbolic computing for
program generationprogram generationprogram generationprogram generation
Input: Input: list of formulas to be programmedlist of formulas to be programmedOutput: Output: corresponding FORTRAN code for corresponding FORTRAN code for
numerical evaluationnumerical evaluation
Sample input:Sample input:
1 subroutine term1d(nvp,kvp,nwp,kwp,nv,kv,nw,kw,jtot,res)2 X[k1](c,d,a,b) X[k2](vp,wp,c,d) Z[k3](v,w,a,b) 3 none4 4xx z1=(-1)**(1+jtot+k1+k2+k3+kapc+kapd+kapw+kapwp) xx z2=d6j(2*jtot,jc,jd,2*k2,jwp,jvp)xx z3=d6j(ja,jb,2*jtot,jw,jv,2*k3)xx z4=d6j(jc,jd,2*jtot,jb,ja,2*k1)5 (e[c,d]-e[vp,wp]) (e[a,b]-e[vp,wp])6 No
( )
1 2 3 '
2 ' ' 3 11 2 3
1 2 3
' '
( 1)
( ) ( ' ' ) ( )
c d a b c dJ k k k j j j jc d w w
abcd k k k w v w v b a
k k k
c d v w a b
J j j J j j J j j
k j j k j j k j j
X cdab X v w cd Z vwab
ε ε ε ε ε ε ε
+ + + + + + + − −
× + − − + −
∑ ∑
( )' 'v wε−
Step 3: Symbolic computing for Step 3: Symbolic computing for Step 3: Symbolic computing for Step 3: Symbolic computing for
program generationprogram generationprogram generationprogram generation
Step 3: Symbolic computing for Step 3: Symbolic computing for Step 3: Symbolic computing for Step 3: Symbolic computing for
program generationprogram generationprogram generationprogram generation
Features: Simple input, essentially just type in a formula.Efficient codes are producedOther constructs can be easily addedSafety features are build inWorks for most formulasFuture plans: interface with formula-writing codes
Step 4: numerical computationStep 4: numerical computationStep 4: numerical computationStep 4: numerical computationStep 4: numerical computationStep 4: numerical computationStep 4: numerical computationStep 4: numerical computation
1. Coupled-cluster method: iterative solution of very large number of equations
2. Relativistic many-body perturbation theory: evaluation of terms such as example considered before
3. Configuration-interaction (CI) method: finding a few eigenvalues of large matrices
4. Combination of coupled-cluster and CI: general super CI code* (in progress)
coupledcoupledcoupledcoupled----cluSter methodcluSter methodcluSter methodcluSter methodcoupledcoupledcoupledcoupled----cluSter methodcluSter methodcluSter methodcluSter method
The excitation coefficients ρ are obtained by the iterative solution of the corresponding equations. Owing to the very large number of the equations the code needs to be very efficient. The inclusion of the triple excitations require new approaches owing to substantially larger memory and computer time requirements.
Cs: a,b = 1s22s22p63s23p63d104s24p64d105s25p6
m,n : finite basis set = (35 ×××× 13) ×××× (35 ×××× 13)19 000 000 equations
Extra index r gives at least a factor of (35 ×××× 13).
bmnaρ
bmnrvaρ
coupledcoupledcoupledcoupled----cluSter methodcluSter methodcluSter methodcluSter methodcoupledcoupledcoupledcoupled----cluSter methodcluSter methodcluSter methodcluSter method
These are really complicated equations !!!
• “Quadruple” term:
• Program has to be exceptionally efficient!
rsmnrs rsabg ρ∑
IndicesIndices mnrsmnrs can be can be ANYANY orbitalsorbitals
Basis set: Basis set: nnmaxmax==3535, , llmaxmax==66
17x17x(17x17x(3535xx1313))44=5 x 10=5 x 101212!!
a,b core(17 shells)
coupledcoupledcoupledcoupled----cluSter methodcluSter methodcluSter methodcluSter methodcoupledcoupledcoupledcoupled----cluSter methodcluSter methodcluSter methodcluSter method
bmnaρ
computational challengeScomputational challengeScomputational challengeScomputational challengeScomputational challengeScomputational challengeScomputational challengeScomputational challengeS
I. Efficiency of the most complicated terms
II. Efficient parallelization
III. Numerical methods
IV. Keeping up with new hardware/software and corresponding compatibility problems
computational challengeScomputational challengeScomputational challengeScomputational challengeScomputational challengeScomputational challengeScomputational challengeScomputational challengeS
1) Efficiency of the most complicated terms:
• Can it be improved further?• Is any time being wasted owing to imperfect
loop ordering?• Automated code optimization?
2) Efficient parallelization, especially for triple excitations
How to do efficient parallelization of iterative procedures with large local files?
rsmnrs rsabg ρ∑
computational challengeScomputational challengeScomputational challengeScomputational challengeScomputational challengeScomputational challengeScomputational challengeScomputational challengeS
3) Efficient parallel codes for configuration interaction method (finding few eigenvalues of large matrices)
4) New approaches for the numerical basis sets (currently using B splines)
5) Resolving convergence issues for the iterative procedure in the cases of large excitations from nearby core shells.
6) Computing on multicore machines
concluSionconcluSionconcluSionconcluSionconcluSionconcluSionconcluSionconcluSion
AtomicClocks
S1/2
P1/2D5/2
Quantumbit
Fundamental Physics
Quantum information
Symbolic and Numeric Symbolic and Numeric Computing for Computing for Atomic CalculationsAtomic Calculations
other theory collaborationS:other theory collaborationS:other theory collaborationS:other theory collaborationS:
Charles Clark and Carl Williams (NIST)Walter Johnson (University of Notre Dame)Ulyana Safronova (University of Nevada-Reno)Michael Kozlov (Petersburg Nuclear Physics Institut e)Victor Flambaum and Vladimir Dzuba (UNSW, Australia )
graduate StudentS:graduate StudentS:graduate StudentS:graduate StudentS:
bindiya arorabindiya arorabindiya arorabindiya arora
rupSi palrupSi palrupSi palrupSi pal
Jenny tchoukovaJenny tchoukovaJenny tchoukovaJenny tchoukova
danSha JiangdanSha JiangdanSha JiangdanSha Jiang
Group web site: http://www.physics.udel.edu/~msafrono
nSf workShopnSf workShopnSf workShopnSf workShop
on cyberon cyberon cyberon cyber----enabled diScovery enabled diScovery enabled diScovery enabled diScovery
and innovationand innovationand innovationand innovation ( cdi )( cdi )( cdi )( cdi )
NSF solicitationhttp://www.nsf.gov/pubs/2007/nsf07603/nsf07603.htm
Workshop pagehttp://www4.ncsu.edu/~kaltofen/CDI_SYMNUM_Itinerary.html