Deep learning jet modifications in heavy-ion collisions
Transcript of Deep learning jet modifications in heavy-ion collisions
June 3, 2021, Jets and their substructure from LHC data, CERN
Deep learning jet modifications in heavy-ioncollisions
JHEP03(2021)206 & arXiv:2106.11271
with Daniel Pablos and Konrad Tywoniuk
Yi-Lun Du
Outline
1 Motivation
2 Deep learning jet energy loss
3 ApplicationSensitivity of jet observables to in-medium modificationJet tomography
4 Conclusion and outlook
Yi-Lun Du Deep learning jet modifications in heavy-ion collisions June 3, 2021 1 / 15
Jets in the medium
14
How can jet modification be quantified?
modified jetjetIdeally…
How do jets from an identical hard process differ in vacuum and in medium?
A-Ap-p
Jasmine Brewer (MIT)
J. Brewer, HP’20
Quark-gluon plasma (QGP) created in heavy ion collision:deconfined phase, hot dense medium
Jets serve as hard probe to the medium properties
Jets are quenched in the medium via parton energy loss
Jasmine Brewer (MIT) 8
Key question: compare A-A jets to which p-p jets?
• Standard answer: match final (reconstructed) !"
RAA =Spectrum in AASpectrum in pp
� E. Epple 7
What we currently know about jet modification
ATLAS collaboration Phys. Lett. B 790 (2019)
ATLAS collaboration Phys. Rev. C 98 (2018)
Jets are: • quenched in central AA collisions• not quenched in MB pp collisions
Jets in the medium appear softerATLAS collaboration PLB 790 (2019) 108
Yi-Lun Du Deep learning jet modifications in heavy-ion collisions June 3, 2021 2 / 15
Jet modifications: ambiguous interpretations
Yaxian MAO Central China Normal University
ALICE Overview (HP2020)
Jet substructure in central Pb-Pb collisions• Soft drop grooming allows to study medium modified parton shower by removing large angle soft radiation
11
• Results are fully corrected for background and detector effects
- No significant modification in zg distribution
- Modification of θg → narrowing jets
Ratio of jet observables distr.between medium and vacuum,BOTH with pjet
T > pcutT
Interplay: jet substructures, e.g., Rg , could– be modified during the passage through the
medium and/or– affect the amount of jet energy loss and then this
jet doesn’t pass the pT cut in the selection, i.e.,selection bias.
Jets produce emissions with smaller Rg inmedium than in vacuum: presumes medium scaledominatesJets with larger Rg in vacuum are moresuppressed in medium: presumes vacuum scaledominatesCan we disentangle these two effects withknowledge of the degree of quenching for eachindividual measured jets?
Yi-Lun Du Deep learning jet modifications in heavy-ion collisions June 3, 2021 3 / 15
Energy loss ratio & Jet selectionsDefining the energy loss ratio
9Daniel Pablos University of Bergen
Vac
Med
Vacuum-like emission
Hypotheticalvacuum-like
emission
Medium inducedemission
<latexit sha1_base64="OF2UJe/9wfLaecmx2PV8SQ2HiIY=">AAAB6nicbVBNS8NAEJ3Ur1q/qh69LBbBU0lE0WNRBI8V7Qe0oWy2m3bpZhN2J0IJ/QlePCji1V/kzX/jts1BWx8MPN6bYWZekEhh0HW/ncLK6tr6RnGztLW9s7tX3j9omjjVjDdYLGPdDqjhUijeQIGStxPNaRRI3gpGN1O/9cS1EbF6xHHC/YgOlAgFo2ilh9te2CtX3Ko7A1kmXk4qkKPeK391+zFLI66QSWpMx3MT9DOqUTDJJ6VuanhC2YgOeMdSRSNu/Gx26oScWKVPwljbUkhm6u+JjEbGjKPAdkYUh2bRm4r/eZ0Uwys/EypJkSs2XxSmkmBMpn+TvtCcoRxbQpkW9lbChlRThjadkg3BW3x5mTTPqt5F1b0/r9Su8ziKcATHcAoeXEIN7qAODWAwgGd4hTdHOi/Ou/Mxby04+cwh/IHz+QMPpI2m</latexit>
Ef
same early, high energy pattern,
up to matching/stopping scale
<latexit sha1_base64="WkCUna+UTfo359QXdqqy98Oh6/I=">AAAB6HicbVDLSgNBEOyNrxhfUY9eBoPgKeyKosegF4+JmAckS5id9CZjZmeXmVkhLPkCLx4U8eonefNvnDwQTSxoKKq66e4KEsG1cd0vJ7eyura+kd8sbG3v7O4V9w8aOk4VwzqLRaxaAdUouMS64UZgK1FIo0BgMxjeTPzmIyrNY3lvRgn6Ee1LHnJGjZVqd91iyS27U5Af4i2SEsxR7RY/O72YpRFKwwTVuu25ifEzqgxnAseFTqoxoWxI+9i2VNIItZ9NDx2TE6v0SBgrW9KQqfp7IqOR1qMosJ0RNQO96E3E/7x2asIrP+MySQ1KNlsUpoKYmEy+Jj2ukBkxsoQyxe2thA2ooszYbAo2hKWXl0njrOxdlN3aealyPY8jD0dwDKfgwSVU4BaqUAcGCE/wAq/Og/PsvDnvs9acM585hD9wPr4Br2GM2w==</latexit>
R
<latexit sha1_base64="WkCUna+UTfo359QXdqqy98Oh6/I=">AAAB6HicbVDLSgNBEOyNrxhfUY9eBoPgKeyKosegF4+JmAckS5id9CZjZmeXmVkhLPkCLx4U8eonefNvnDwQTSxoKKq66e4KEsG1cd0vJ7eyura+kd8sbG3v7O4V9w8aOk4VwzqLRaxaAdUouMS64UZgK1FIo0BgMxjeTPzmIyrNY3lvRgn6Ee1LHnJGjZVqd91iyS27U5Af4i2SEsxR7RY/O72YpRFKwwTVuu25ifEzqgxnAseFTqoxoWxI+9i2VNIItZ9NDx2TE6v0SBgrW9KQqfp7IqOR1qMosJ0RNQO96E3E/7x2asIrP+MySQ1KNlsUpoKYmEy+Jj2ukBkxsoQyxe2thA2ooszYbAo2hKWXl0njrOxdlN3aealyPY8jD0dwDKfgwSVU4BaqUAcGCE/wAq/Og/PsvDnvs9acM585hD9wPr4Br2GM2w==</latexit>
R
<latexit sha1_base64="8yIx1DVd7N0LFhlo1TkmcdQfAzo=">AAAB6nicbVBNS8NAEJ3Ur1q/qh69LBbBU0lE0WNRBI8V7Qe0oWy2k3bpZhN2N0IJ/QlePCji1V/kzX/jts1BWx8MPN6bYWZekAiujet+O4WV1bX1jeJmaWt7Z3evvH/Q1HGqGDZYLGLVDqhGwSU2DDcC24lCGgUCW8HoZuq3nlBpHstHM07Qj+hA8pAzaqz0cNvjvXLFrbozkGXi5aQCOeq98le3H7M0QmmYoFp3PDcxfkaV4UzgpNRNNSaUjegAO5ZKGqH2s9mpE3JilT4JY2VLGjJTf09kNNJ6HAW2M6JmqBe9qfif10lNeOVnXCapQcnmi8JUEBOT6d+kzxUyI8aWUKa4vZWwIVWUGZtOyYbgLb68TJpnVe+i6t6fV2rXeRxFOIJjOAUPLqEGd1CHBjAYwDO8wpsjnBfn3fmYtxacfOYQ/sD5/AEUMI2p</latexit>
Ei
The amount of energy and momentum lost by the energetic parton, as described by
Eq. (2.1), exactly corresponds to the amount of energy and momentum flowing into the
QGP hydrodynamic modes [70]. This will generate a wake that is correlated with the direc-
tion of the jet [78], whose contribution to the experimentally observable jet properties has
to be taken into account. The hybrid model provides an estimate of the wake contribution
to the final hadron spectrum by performing an expansion of the Cooper-Frye formula at
the perturbed freeze-out hypersurface, which yields [79]
Ed�N
d3p=
1
32⇡
mT
T 5cosh(y � yj) exp
h�mT
Tcosh(y � yj)
i
⇥(
pT�PT cos(�� �j) +1
3mT �MT cosh(y � yj)
),
(2.5)
where pT , mT , � and y are the transverse momentum, transverse mass, azimuthal angle
and rapidity of the emitted thermal particles and where �PT and �MT = �E/ cosh yj are
the transverse momentum and transverse mass transferred from the jet, with azimuthal
angle and rapidity �j and yj , respectively. The distribution in Eq. (2.5) has been obtained
by considering that the background behaves as a Bjorken flow, which only has longitudinal
expansion. Generalizing it to the case in which there is transverse expansion can modify
such distribution, depending on the orientation of the jet with respect to the background
radial flow components [80–82]. The consequences of these observations will be explored
in the near future.
The partons that do not completely hydrodynamize are hadronized using the Lund
string model included in PYTHIA8. The contributions from the hadrons of the wake,
together with the fragmented hadrons, ensure event-by-event energy-momentum conserva-
tion.3
2.2 Jet energy loss ratio �jh and traversed path-length L
The main goal of this work is to determine, on a jet-by-jet basis, the amount of energy
loss, quantified through the variable
�jh ⌘Eh
f
Ehi
, (2.6)
su↵ered by jets due to the propagation through a hot and dense QCD medium. Here,
the subscript “jh” refers to the energy of the jet measured at hadronic level. These jets
3The distribution in Eq. (2.5) can become negative, most notably in the direction opposite to the jet in
the transverse plane. This reflects the absence of soft particles in such region of phase space compared to
an unperturbed QGP background as a result to the boost experienced by the fluid cell due to the injection
of momentum from the jet. In the present work we will ignore such negative contributions, since they
would show up as negative energy pixels in the jet images used in Section 3.1 (one would need to devise a
procedure to cancel out such negative contributions using particles from a real background which are close
in momentum and configuration space, such as in [79], which we leave for future work). It has been shown
that their contribution to jet observables with relatively small jet radius, such as the one used in the present
work, R = 0.4, is almost negligible [83], which guarantees that none of our conclusions will be a↵ected by
the omission of such contribution. A study of jets with a larger radius will be done in future publications.
– 6 –
FESIES
Study jet observables for jets that belong to 2 differentquenching classes:
– Unquenched class: χjh > 0.9.– Quenched class: χjh < 0.9.
pp jets: pT > 200 GeVPbPb jets:
– Final Energy Selection (FES): impose pT cut onfinal energy pT > 200 GeV→ Steeply falling energyloss dist. Biased by little quenched samples!
– Initial Energy Selection (IES): impose pT cut oninitial energy via χjh, pT/χjh > 200 GeV & pT > 100GeV→ More support of fairly quenched jets in thequenched class. More distinguishable!
Yi-Lun Du Deep learning jet modifications in heavy-ion collisions June 3, 2021 4 / 15
Hybrid model
Hybrid Model
11Daniel Pablos University of Bergen
free parameterO(1)
Strongly coupled energy loss(hydrodynamization rate)
Hadrons from the hydro.wake (medium response)
PYTHIA8 down to hadro. scale(formation time argumentfor spacetime picture)
Can also studyfinite resolution effects(not done in present work)
Casalderrey-Solana, Gulhan, Milhano, DP, Rajagopal JHEP ‘15,‘16,‘17
Hulcher, DP, Rajagopal JHEP ‘18
PYTHIA8 down to hadronization scale
Strongly coupled energy loss at every stage
Hadrons from the hydro. wake (medium response)Casalderrey-Solana, Gulhan, Milhano, Daniel Pablos, Rajagopal JHEP ’15,’16,’17
Vacuum jets using p̂T ,min = 50GeV, with oversampling powerp4
T .
PbPb collisions in 0-5%centrality at
√s = 5.02 ATeV.
Reconstructed jets with anti-kT ,R = 0.4, required to be |η| < 2and pjet
T > 100 GeV.
∼ 250,000 jets. 80% for trainingand 20% for validation.
Yi-Lun Du Deep learning jet modifications in heavy-ion collisions June 3, 2021 5 / 15
CNN Prediction & Interpretability
χjh
Jet image33x33
16features17x17
16features17x17
32features
9x9flattened
2592fc
128outputlayer
8x8 conv, 16bn, PReLu
dropout(0.2)avgpool(2x2)
7x7x16 conv, 16bn, PReLu
dropout(0.2)
6x6x16 conv, 32bn, PReLu
dropout(0.2)avgpool(2x2)
bn, PReLudropout(0.5)
0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0True jh
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1.0
Pred
icted
jh
True jh versus predicted jh
0.00
0.05
0.10
0.15
0.20
0.25
0.30
0.35
0.40
0.4
0.3
0.2
0.1
0.0
0.1
0.2
0.3
0.4
Tran
slate
d Az
imut
hal A
ngle
Average of normalized jet image, 0.25< jh < 0.50
Average of normalized jet image, 0.50< jh < 0.60
Average of normalized jet image, 0.60< jh < 0.70
0.4 0.2 0.0 0.2 0.4Translated Pseudorapidity
0.4
0.3
0.2
0.1
0.0
0.1
0.2
0.3
0.4
Tran
slate
d Az
imut
hal A
ngle
Average of normalized jet image, 0.70< jh < 0.85
0.4 0.2 0.0 0.2 0.4Translated Pseudorapidity
Average of normalized jet image, 0.85< jh < 1
0.4 0.2 0.0 0.2 0.4Translated Pseudorapidity
Correlation of jh with per-pixel of jet image
10 6
10 5
10 4
10 3
10 2
10 1
0.3
0.2
0.1
0.0
0.1
0.2
0.3
Jet quenching increases the numberof soft particles at large anglesJet shape can capture the main feature
Yi-Lun Du Deep learning jet modifications in heavy-ion collisions June 3, 2021 6 / 15
Jet radius, RgRg ratio between PbPb and pp jets
FES: Selection bias towards jets withsmaller Rg , originated by pT cut.
IES:
– Unquenched class: still biased due toχjh cut: to belong to this class, a jet hadbetter to be with smaller Rg , comparedwith all pp jets.
– Quenched class presents featuresrelated to energy loss, compared withunquenched class: jet quenching leadsto enhancement of large Rg .
FES IES
Yi-Lun Du Deep learning jet modifications in heavy-ion collisions June 3, 2021 7 / 15
Soft Drop multiplicity, nSDnSD ratio between PbPb and pp jets
FES: Selection bias towards jets withfewer nSD, originated by pT cut.
IES:
– Unquenched class: still biased due toχjh cut: to belong to this class, a jet hadbetter to be with fewer nSD, comparedwith all pp jets.
– Quenched class presents featuresrelated to energy loss, compared withunquenched class: jet quenching leadsto enhancement of large nSD.
FES IES
Yi-Lun Du Deep learning jet modifications in heavy-ion collisions June 3, 2021 8 / 15
Groomed momentum sharing fraction, zgzg ratio between PbPb and pp jets
FES: No selection bias observed. Scaleof emission isn’t strongly dependent onsplitting fraction zg .
IES:
– Quenched class presents featuresrelated to energy loss, compared withunquenched class: jet quenching leadsto enhancement of smaller zg subjets.
FES IES
Y.-L. Du, D. Pablos, K. Tywoniuk, JHEP03(2021)206
Yi-Lun Du Deep learning jet modifications in heavy-ion collisions June 3, 2021 9 / 15
Applications: Jet tomography, length VS χjh
0 2 4 6 8 10 12 14 16L (fm)
0.00
0.05
0.10
0.15
0.20
0.25
0.30
0.35
0.40
0.45
Norm
alize
d to
Uni
ty
Histogram for L w/weights0.95< p
jh<10.85< p
jh < 0.950.75< p
jh < 0.850.60< p
jh < 0.750.25< p
jh<0.60
Due to the strong correlationbetween L and χjh, selecting
jets with different χjh willnaturally select jets that
traversed different L.→ Great potential to maketomographic application!
5
0
5
10
y (f
m)
0.25< pjh < 0.60 0.60< p
jh < 0.75 0.75< pjh < 0.85
10 5 0 5x (fm)
5
0
5
10
y (f
m)
0.85< pjh < 0.95
10 5 0 5x (fm)
0.95< pjh < 1
10 5 0 5x (fm)
Inclusive in pjh
0.005
0.010
0.015
0.020
0.025
0.030
Yi-Lun Du Deep learning jet modifications in heavy-ion collisions June 3, 2021 10 / 15
Applications: creation points & orientation
5
0
5
10
y (f
m)
FES
Creation points for centrality 0-5%FES-Glauber
5
0
5
10
y (f
m)
IES IES-Glauber
10 5 0 5x (fm)
5
0
5
10
y (f
m)
IES, Predicted
10 5 0 5x (fm)
IES, Predicted-Glauber
0.0025
0.0050
0.0075
0.0100
0.0125
0.0150
0.0175
0.0200
0.006
0.004
0.002
0.000
0.002
0.004
0.006
v2 =
⟨p2
x−p2y
p2x+p2
y
⟩
0 10 20 30 40 50 60 70Centrality (%)
0.01
0.00
0.01
0.02
0.03
0.04
0.05
v 2
v2 VS Centrality for R = 0.4 @ sNN = 5.02 TeVFESIESIES, Predicted
IES “removes” final state interactions (selectionbias), since we record “all” jets.IES provides access to the genuine jet creationpoint distribution and initial orientation.
Y.-L. Du, D. Pablos, K. Tywoniuk, arXiv: 2106.11271
Yi-Lun Du Deep learning jet modifications in heavy-ion collisions June 3, 2021 11 / 15
Conclusion and outlookCNN can extract energy loss jet-by-jet from jet image with good performanceProcedure generalisable to many jet quenching modelsJet shape contains significant predictive power: angular distribution of soft particlesMitigate selection bias and reveal medium effects on various jet observablesOpen opportunity to make tomographic study
– Generalizability to other MC quenching models?– Applicability to more realistic environment: fluctuating background?– Better performance from other state-of-the-art neural networks?– Extract traversed length with better precision?– Unfold jet initial properties apart from jet energy?
Yi-Lun Du Deep learning jet modifications in heavy-ion collisions June 3, 2021 12 / 15
Backup: Groomed jet image VS χjh
0.4
0.3
0.2
0.1
0.0
0.1
0.2
0.3
0.4
Tran
slate
d Az
imut
hal A
ngle
Average of normalized groomed jet image, 0.25< jh < 0.50
Average of normalized groomed jet image, 0.50< jh < 0.60
Average of normalized groomed jet image, 0.60< jh < 0.70
0.4 0.2 0.0 0.2 0.4Translated Pseudorapidity
0.4
0.3
0.2
0.1
0.0
0.1
0.2
0.3
0.4
Tran
slate
d Az
imut
hal A
ngle
Average of normalized groomed jet image, 0.70< jh < 0.85
0.4 0.2 0.0 0.2 0.4Translated Pseudorapidity
Average of normalized groomed jet image, 0.85< jh < 1
0.4 0.2 0.0 0.2 0.4Translated Pseudorapidity
Correlation of jh with per-pixel of groomed jet image
10 6
10 5
10 4
10 3
10 2
10 1
0.3
0.2
0.1
0.0
0.1
0.2
0.3
Input Output Network LossGroomed jet image χjh CNN 0.0065
Jet image above 1 GeV χjh CNN 0.0042Jet image above 2 GeV χjh CNN 0.0066
Removing soft particles (atlarge angles) reducesperformance
Generalizability challenges:
– Large fluctuatingbackground
– Hardonization modelling– Other quenching models
Yi-Lun Du Deep learning jet modifications in heavy-ion collisions June 3, 2021 13 / 15
Backup: Prediction performance with FCNNInput (size) Output Network Loss
FF (10) χjh FCNN 0.0058Jet shape (8) χjh FCNN 0.0033
FF, jet shape (18) χjh FCNN 0.0032FF, jet shape, features (25) χjh FCNN 0.0028
Jet image & FF, jet shape, features (25) χjh API: CNN&FCNN 0.0028
Jet shape outperforms jet FF.
Motivates construction from jetshape by 17-parameter fitting:
– Still a bit worse than CNN
Jet observables recover theperformance by jet image withequivalent predictive power:interpretability!
0.05 0.10 0.15 0.20 0.25 0.30 0.35r
10 2
10 1
(r)(P
bPb)
Jet shape0.95< jh<10.85< jh<0.950.75< jh<0.850.60< jh<0.750.25< jh<0.60Medium
10 1 100
z
10 2
10 1
100
101
D(z)
(PbP
b)
FF
0.95< jh<10.85< jh<0.950.75< jh<0.850.60< jh<0.750.25< jh<0.60Medium
0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0True jh
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1.0
Fitte
d js jh
True jh versus fitted jsjh by jet shape
0.00
0.05
0.10
0.15
0.20
0.25
0.30
0.35
0.40
χjsjh =
∑i
(pTipT
)αirβii + γ
Yi-Lun Du Deep learning jet modifications in heavy-ion collisions June 3, 2021 14 / 15
Backup: Jet tomography with χjh & v2
v2 =p2
x−p2y
p2x+p2
y
Top row: In-plane jets(v2 > 0) going left (px < 0)and right (px > 0)Bottom row: Out-of-planejets (v2 < 0) going up(py > 0) and down (py < 0)To get very quenched, jetshave to travel longer inmedium. So v2 & px ,y arehelpful for jet tomography.
x (fm)
10.07.55.0
2.50.0
2.55.0
7.510.0
y (fm)
10.07.5
5.02.5
0.02.5
5.07.510.0
0.02
0.00
0.02
0.04
0.06
0.08
0.10
0.25< p<0.75
x (fm)
10.07.55.0
2.50.0
2.55.0
7.510.0
y (fm)
10.07.5
5.02.5
0.02.5
5.07.510.0
0.02
0.00
0.02
0.04
0.06
0.08
0.10
Creation points density for centrality 30-40%, R = 0.4 @ sNN = 5.02 TeV, FES, pT > 100 GeV
In-plane
0.75< p<0.95
x (fm)
10.07.55.0
2.50.0
2.55.0
7.510.0
y (fm)
10.07.5
5.02.5
0.02.5
5.07.510.0
0.02
0.00
0.02
0.04
0.06
0.08
0.10
0.95< p<1.00
x (fm)
10.07.55.0
2.50.0
2.55.0
7.510.0
y (fm)
10.07.5
5.02.5
0.02.5
5.07.510.0
0.02
0.00
0.02
0.04
0.06
0.08
0.10
Inclusive in p
x (fm)
10.07.55.0
2.50.0
2.55.0
7.510.0
y (fm)
10.07.5
5.02.5
0.02.5
5.07.510.0
0.02
0.00
0.02
0.04
0.06
0.08
0.10
0.25< p<0.75
x (fm)
10.07.55.0
2.50.0
2.55.0
7.510.0
y (fm)
10.07.5
5.02.5
0.02.5
5.07.510.0
0.02
0.00
0.02
0.04
0.06
0.08
0.10
Out-of-plane
0.75< p<0.95
x (fm)
10.07.55.0
2.50.0
2.55.0
7.510.0
y (fm)
10.07.5
5.02.5
0.02.5
5.07.510.0
0.02
0.00
0.02
0.04
0.06
0.08
0.10
0.95< p<1.00
x (fm)
10.07.55.0
2.50.0
2.55.0
7.510.0
y (fm)
10.07.5
5.02.5
0.02.5
5.07.510.0
0.02
0.00
0.02
0.04
0.06
0.08
0.10
Inclusive in p
0.000
0.006
0.012
0.018
0.024
0.030
0.036
0.042
Yi-Lun Du Deep learning jet modifications in heavy-ion collisions June 3, 2021 15 / 15