machine learning planning exercise: theory applications ...ml for qcd global analysis - proposal for...
TRANSCRIPT
1 / 2
Machine Learning Planning Exercise:Theory Applications
Nobuo SatoODU/JLab
Computing Round Table (2019)JLab, 2019
Zooming in at the femtometerscale using JLab12
2 / 2
e
e′
P
Zooming in at the femtometerscale using JLab12
2 / 2
e
e′
PInverse problem
(computing)
Factorization(theory)
Quantum probability distributions in the nucleon
3 / 2
dσ
dx dy dΨ dz dφh dP2hT
∼18∑i=1
Fi(x, z,Q2, P 2hT )βi
Fi Standard label βiF1 FUU,T 1F2 FUU,L ε
F3 FLL S||λe√
1− ε2
F4 Fsin(φh+φS)UT |~S⊥|ε sin(φh + φS)
F5 Fsin(φh−φS)UT,T |~S⊥|sin(φh − φS)
F6 Fsin(φh−φS)UT,L |~S⊥|ε sin(φh − φS)
F7 F cos 2φhUU ε cos(2φh)
F8 Fsin(3φh−ψS)UT |~S⊥|ε sin(3φh − φS)
F9 Fcos(φh−φS)LT |~S⊥|λe
√1− ε2 cos(φh − φS)
F10 F sin 2φhUL S||ε sin(2φh)
F11 F cosφSLT |~S⊥|λe
√2ε(1− ε) cosφS
F12 F cosφhLL S||λe
√2ε(1− ε) cosφh
F13 Fcos(2φh−φS)LT |~S⊥|λe
√2ε(1− ε) cos(2φh − φS)
F14 F sinφhUL S||
√2ε(1 + ε) sinφh
F15 F sinφhLU λe
√2ε(1− ε) sinφh
F16 F cosφhUU
√2ε(1 + ε) cosφh
F17 F sinφSUT |~S⊥|
√2ε(1 + ε) sinφS
F18 Fsin(2φh−φS)UT |~S⊥|
√2ε(1 + ε) sin(2φh − φS)
Name Symbol meaningupol. PDF f q1 U. pol. quarks in U. pol. nucleonpol. PDF gq1 L. pol. quarks in L. pol. nucleon
Transversity hq1 T. pol. quarks in T. pol. nucleonSivers f
⊥(1)q1T U. pol. quarks in T. pol. nucleon
Boer-Mulders h⊥(1)q1 T. pol. quarks in U. pol. nucleon
......
...FF Dq
1 U. pol. quarks to U. pol. hadronCollins H
⊥(1)q1 T. pol. quarks to U. pol. hadron
......
...
QCD global analysis in a nutshell
4 / 2
Expe
rimen
ts lN
NN
ll
QCD global analysis in a nutshell
4 / 2
Expe
rimen
ts lN
NN
ll
DIS
SIDIS
DY
SIA
QCD global analysis in a nutshell
4 / 2
Expe
rimen
ts lN
NN
ll
DIS
SIDIS
DY
SIA
Theory=
QC
D
FF
DIS
SIDIS
DY
SIA
QCD global analysis in a nutshell
4 / 2
Expe
rimen
ts lN
NN
ll
DIS
SIDIS
DY
SIA
Theory=
QC
D
FF
DIS
SIDIS
DY
SIA
Parameters
QCD global analysis in a nutshell
4 / 2
Expe
rimen
ts lN
NN
ll
DIS
SIDIS
DY
SIA
Theory=
QC
D
FF
DIS
SIDIS
DY
SIA
Parameters
Exp ?= Thy
No“tuning”
Yes
QCD global analysis in a nutshell
4 / 2
Expe
rimen
ts lN
NN
ll
DIS
SIDIS
DY
SIA
Theory=
QC
D
FF
DIS
SIDIS
DY
SIA
Parameters
Exp ?= Thy
No“tuning”
Yes Physics interpretation
QCD global analysis in a nutshell
4 / 2
Expe
rimen
ts lN
NN
ll
DIS
SIDIS
DY
SIA
Theory=
QC
D
FF
DIS
SIDIS
DY
SIA
Parameters
Exp ?= Thy
No“tuning”
Yes Physics interpretationOptimize measurements
ML for global analysis
5 / 2
Input Output
Theory
Ω
dσ/d
Ω
ParameterSpace
ObservableSpace
Ω
dσ/d
Ω
NeuralN
ets
ParameterSpace
ObservableSpace
Ω
dσ/d
Ω
NeuralN
ets
ExperimentalData Visualization
Application
Training
Theory: PDFs, FFs, TMDs,GPDs, GTMDs, Wignerdistributions
ML: parameter space→ observable space
Multi-disciplinary:
o QCD scientists: JLab,Argonne, Temple
o Comp. scientists: ODU,Davidson College
Proposal for CNF
Empirically Trained Hadronic Event Regenerator (ETHER)
6 / 2
Nature
Events:vertex level
0.0 0.2 0.4 0.6 0.8 1.00
50
100
150
200
250
LDRD19:JLab/ODU/Davidson
Empirically Trained Hadronic Event Regenerator (ETHER)
6 / 2
Nature
Events:vertex level
0.0 0.2 0.4 0.6 0.8 1.00
50
100
150
200
250
Experimentaldetector
Events:detector level
0.0 0.2 0.4 0.6 0.8 1.00
100
200
300
400distortion
LDRD19:JLab/ODU/Davidson
Empirically Trained Hadronic Event Regenerator (ETHER)
6 / 2
Nature
Events:vertex level
0.0 0.2 0.4 0.6 0.8 1.00
50
100
150
200
250
Experimentaldetector
Events:detector level
0.0 0.2 0.4 0.6 0.8 1.00
100
200
300
400distortion
Inverse problem → solutions are not unique
→ model dependentLDRD19:JLab/ODU/Davidson
Empirically Trained Hadronic Event Regenerator (ETHER)
6 / 2
Nature
Events:vertex level
0.0 0.2 0.4 0.6 0.8 1.00
50
100
150
200
250
Experimentaldetector
Events:detector level
0.0 0.2 0.4 0.6 0.8 1.00
100
200
300
400distortion
ETHER
Events:vertex level
0.0 0.2 0.4 0.6 0.8 1.00
50
100
150
200
250
LDRD19:JLab/ODU/Davidson
Empirically Trained Hadronic Event Regenerator (ETHER)
6 / 2
Nature
Events:vertex level
0.0 0.2 0.4 0.6 0.8 1.00
50
100
150
200
250
Experimentaldetector
Events:detector level
0.0 0.2 0.4 0.6 0.8 1.00
100
200
300
400distortion
ETHER
Events:vertex level
0.0 0.2 0.4 0.6 0.8 1.00
50
100
150
200
250
Detectorsimulator
neural netdetector
LDRD19:JLab/ODU/Davidson
Empirically Trained Hadronic Event Regenerator (ETHER)
6 / 2
Nature
Events:vertex level
0.0 0.2 0.4 0.6 0.8 1.00
50
100
150
200
250
Experimentaldetector
Events:detector level
0.0 0.2 0.4 0.6 0.8 1.00
100
200
300
400distortion
ETHER
Events:vertex level
0.0 0.2 0.4 0.6 0.8 1.00
50
100
150
200
250
Detectorsimulator
neural netdetector
Events:detector level
0.0 0.2 0.4 0.6 0.8 1.00
100
200
300
400distortionLDRD19:JLab/ODU/Davidson
Empirically Trained Hadronic Event Regenerator (ETHER)
6 / 2
Nature
Events:vertex level
0.0 0.2 0.4 0.6 0.8 1.00
50
100
150
200
250
Experimentaldetector
Events:detector level
0.0 0.2 0.4 0.6 0.8 1.00
100
200
300
400distortion
ETHER
Events:vertex level
0.0 0.2 0.4 0.6 0.8 1.00
50
100
150
200
250
Detectorsimulator
neural netdetector
Events:detector level
0.0 0.2 0.4 0.6 0.8 1.00
100
200
300
400distortion
Training
LDRD19:JLab/ODU/Davidson
Empirically Trained Hadronic Event Regenerator (ETHER)
6 / 2
Nature
Events:vertex level
0.0 0.2 0.4 0.6 0.8 1.00
50
100
150
200
250
Experimentaldetector
Events:detector level
0.0 0.2 0.4 0.6 0.8 1.00
100
200
300
400distortion
ETHER
Events:vertex level
0.0 0.2 0.4 0.6 0.8 1.00
50
100
150
200
250
Detectorsimulator
neural netdetector
Events:detector level
0.0 0.2 0.4 0.6 0.8 1.00
100
200
300
400distortion
Training
datacom
pression
LDRD19:JLab/ODU/Davidson
Summary and outlook
7 / 2
ML for QCD global analysis - proposal for CNF
o Multi-disciplinary → QCD scientists, computer scientists
o Next generation of QCD analysis tools → boost scientific research
ML based MCEG (ETHER) - LDRD19
o Data compactification tool
o MCEG free of theory assumptions at the femtometer scale