HEP.TrkX Charged Particle Tracking using Graph Neural Networks · HEP.TrkX Charged Particle...
Transcript of HEP.TrkX Charged Particle Tracking using Graph Neural Networks · HEP.TrkX Charged Particle...
![Page 1: HEP.TrkX Charged Particle Tracking using Graph Neural Networks · HEP.TrkX Charged Particle Tracking using Graph Neural Networks. ... Long Short Term Memory - LSTM Breakthrough in](https://reader034.fdocuments.us/reader034/viewer/2022042220/5ec67716e690c36f2e14497e/html5/thumbnails/1.jpg)
ACAT, Saas-Fee, 10-15 March 2019
Jean-Roch Vlimant for the HEP.TrkX projectSpecial credits to
Xiangyang Ju and Alexander Zlokapa
HEP.TrkX Charged Particle Tracking
using Graph Neural Networks
![Page 2: HEP.TrkX Charged Particle Tracking using Graph Neural Networks · HEP.TrkX Charged Particle Tracking using Graph Neural Networks. ... Long Short Term Memory - LSTM Breakthrough in](https://reader034.fdocuments.us/reader034/viewer/2022042220/5ec67716e690c36f2e14497e/html5/thumbnails/2.jpg)
03/10/19 2J.-R. Vlimant
HEP.TrkX Project➢ Pilot project funded by DOE ASCR and COMP HEP➢ Part of HEP CCE➢ Mission
Explore deep learning techniques for track formation
➢ People LBL : Paolo Calafiura, Steve Farrell, Mayur Mudigonda,
Prabhat FNAL : Giuseppe Cerati, Lindsey Gray, Jim Kowalkowski,
Panagiotis Spentzouris, Aristeidis Tsaris Caltech : Dustin Anderson, Josh Bendavid, Pietro Perona,
Maria Spiropulu, Jean-Roch Vlimant, Stephan Zheng
➢ All material available under https://heptrkx.github.io/
![Page 3: HEP.TrkX Charged Particle Tracking using Graph Neural Networks · HEP.TrkX Charged Particle Tracking using Graph Neural Networks. ... Long Short Term Memory - LSTM Breakthrough in](https://reader034.fdocuments.us/reader034/viewer/2022042220/5ec67716e690c36f2e14497e/html5/thumbnails/3.jpg)
03/10/19 3J.-R. Vlimant
Outline
➢ The challenge of Charged Particle Tracking
➢ Forewords on tracking with ML
➢ Dataset and graph neural network models
➢ Results and outlooks
![Page 4: HEP.TrkX Charged Particle Tracking using Graph Neural Networks · HEP.TrkX Charged Particle Tracking using Graph Neural Networks. ... Long Short Term Memory - LSTM Breakthrough in](https://reader034.fdocuments.us/reader034/viewer/2022042220/5ec67716e690c36f2e14497e/html5/thumbnails/4.jpg)
03/10/19 4J.-R. Vlimant
Tracking in a Nutshell
Seeding Kalman Filter
● Particle trajectory bended in asolenoid magnetic field
● Curvature is a proxy tomomentum
● Particle ionize silicon pixel and strip throughout several concentric layers
● Thousands of sparse hits● Lots of hit pollution from low
momentum, secondary particles
● Explosion of hit combinatorics in both seeding and stepping pattern recognition● Highly time consuming task in extracting physics content from LHC data
![Page 5: HEP.TrkX Charged Particle Tracking using Graph Neural Networks · HEP.TrkX Charged Particle Tracking using Graph Neural Networks. ... Long Short Term Memory - LSTM Breakthrough in](https://reader034.fdocuments.us/reader034/viewer/2022042220/5ec67716e690c36f2e14497e/html5/thumbnails/5.jpg)
03/10/19 5J.-R. Vlimant
Complexity and Ambiguity
Shown trajectories are reconstructed objects
The future holds much more hits
![Page 6: HEP.TrkX Charged Particle Tracking using Graph Neural Networks · HEP.TrkX Charged Particle Tracking using Graph Neural Networks. ... Long Short Term Memory - LSTM Breakthrough in](https://reader034.fdocuments.us/reader034/viewer/2022042220/5ec67716e690c36f2e14497e/html5/thumbnails/6.jpg)
03/10/19 6J.-R. Vlimant
High Luminosity LHCThe Challenge
![Page 7: HEP.TrkX Charged Particle Tracking using Graph Neural Networks · HEP.TrkX Charged Particle Tracking using Graph Neural Networks. ... Long Short Term Memory - LSTM Breakthrough in](https://reader034.fdocuments.us/reader034/viewer/2022042220/5ec67716e690c36f2e14497e/html5/thumbnails/7.jpg)
03/10/19 7J.-R. Vlimant
Cost of Tracking● CPU time consumption in HL-LHC era surpasses computing budget
➔ Need for faster algorithms● Charged particle track reconstruction is one of the most CPU consuming
task in event reconstruction➔ Optimizations mostly saturated
● Large fraction of CPU required in the HLT. Cannot perform trackinginclusively
➔ Approximation possible in the trigger
@HL-LHC <μ> >200
![Page 8: HEP.TrkX Charged Particle Tracking using Graph Neural Networks · HEP.TrkX Charged Particle Tracking using Graph Neural Networks. ... Long Short Term Memory - LSTM Breakthrough in](https://reader034.fdocuments.us/reader034/viewer/2022042220/5ec67716e690c36f2e14497e/html5/thumbnails/8.jpg)
03/10/19 8J.-R. Vlimant
Fast Hardware Tracking● Track trigger implementation for Trigger
upgrades development on-going● Several approaches investigated● Dedicated hardware is the key to fast
computation.● Not applicable for offline processing unless
through adopting heterogeneous computing.Tracklets
Hough
Transform
Kalman
Filter
![Page 9: HEP.TrkX Charged Particle Tracking using Graph Neural Networks · HEP.TrkX Charged Particle Tracking using Graph Neural Networks. ... Long Short Term Memory - LSTM Breakthrough in](https://reader034.fdocuments.us/reader034/viewer/2022042220/5ec67716e690c36f2e14497e/html5/thumbnails/9.jpg)
03/10/19 9J.-R. Vlimant
Motivations
Current algorithms for tracking are highlyperformant physics-wise and scale badly
computation-wise
Faster implementations are possible withdedicated hardware
Go back to the blackboard for new approaches
![Page 10: HEP.TrkX Charged Particle Tracking using Graph Neural Networks · HEP.TrkX Charged Particle Tracking using Graph Neural Networks. ... Long Short Term Memory - LSTM Breakthrough in](https://reader034.fdocuments.us/reader034/viewer/2022042220/5ec67716e690c36f2e14497e/html5/thumbnails/10.jpg)
03/10/19 10J.-R. Vlimant
HEP.TrkX Approaches
Seq-to-seq track finding
https://tinyurl.com/y87saehf
https://tinyurl.com/yb3v93y9
End-to-end hit assignment
Track following with RNN https://heptrkx.github.io/
![Page 11: HEP.TrkX Charged Particle Tracking using Graph Neural Networks · HEP.TrkX Charged Particle Tracking using Graph Neural Networks. ... Long Short Term Memory - LSTM Breakthrough in](https://reader034.fdocuments.us/reader034/viewer/2022042220/5ec67716e690c36f2e14497e/html5/thumbnails/11.jpg)
03/10/19 11J.-R. Vlimant
Data Representation
![Page 12: HEP.TrkX Charged Particle Tracking using Graph Neural Networks · HEP.TrkX Charged Particle Tracking using Graph Neural Networks. ... Long Short Term Memory - LSTM Breakthrough in](https://reader034.fdocuments.us/reader034/viewer/2022042220/5ec67716e690c36f2e14497e/html5/thumbnails/12.jpg)
03/10/19 12J.-R. Vlimant
Charged Particle Tracking Dataset
https://www.kaggle.com/c/trackml-particle-identificationhttps://competitions.codalab.org/competitions/20112
● This work uses the publicdataset of the TrackMLParticle Tracking Challenge(Kaggle, codalab).
● Simulating the denseenvironment expected forHL-HLC. Average of 200proton-proton interaction perbunch crossing.
![Page 13: HEP.TrkX Charged Particle Tracking using Graph Neural Networks · HEP.TrkX Charged Particle Tracking using Graph Neural Networks. ... Long Short Term Memory - LSTM Breakthrough in](https://reader034.fdocuments.us/reader034/viewer/2022042220/5ec67716e690c36f2e14497e/html5/thumbnails/13.jpg)
03/10/19 13J.-R. Vlimant
Tracker Hit Graph
Directed graph constructed➢ One tracker hit per node➢ Direct edge inside-out
![Page 14: HEP.TrkX Charged Particle Tracking using Graph Neural Networks · HEP.TrkX Charged Particle Tracking using Graph Neural Networks. ... Long Short Term Memory - LSTM Breakthrough in](https://reader034.fdocuments.us/reader034/viewer/2022042220/5ec67716e690c36f2e14497e/html5/thumbnails/14.jpg)
03/10/19 14J.-R. Vlimant
Edge Classification withGraph Neural Network
![Page 15: HEP.TrkX Charged Particle Tracking using Graph Neural Networks · HEP.TrkX Charged Particle Tracking using Graph Neural Networks. ... Long Short Term Memory - LSTM Breakthrough in](https://reader034.fdocuments.us/reader034/viewer/2022042220/5ec67716e690c36f2e14497e/html5/thumbnails/15.jpg)
03/10/19 15J.-R. Vlimant
Node & Edge Representations
Node representation
Tracker hit feature
Edge representation
Vector
Edge ScoreEdge Score
Latent Space
Output
Input
Latent edge representation taken to be the classification scoreinstead of some latent vector representation
![Page 16: HEP.TrkX Charged Particle Tracking using Graph Neural Networks · HEP.TrkX Charged Particle Tracking using Graph Neural Networks. ... Long Short Term Memory - LSTM Breakthrough in](https://reader034.fdocuments.us/reader034/viewer/2022042220/5ec67716e690c36f2e14497e/html5/thumbnails/16.jpg)
03/10/19 16J.-R. Vlimant
Neural Networks● Input Network
➢ Transforms from hit features (r,φ , z) to the node latentrepresentation (N for 8 to 128)
Dense : 3→...→N
● Edge Network➢ Predicts an edge weight from the node latent
representation at both ends Dense : N+N→...→1
● Node Network➢ Predicts a node latent representation from the current
node representation, weighted sum of node latentrepresentation from incoming edge, and weighted sum
Dense : N+N+N→...→N
![Page 17: HEP.TrkX Charged Particle Tracking using Graph Neural Networks · HEP.TrkX Charged Particle Tracking using Graph Neural Networks. ... Long Short Term Memory - LSTM Breakthrough in](https://reader034.fdocuments.us/reader034/viewer/2022042220/5ec67716e690c36f2e14497e/html5/thumbnails/17.jpg)
03/10/19 17J.-R. Vlimant
Edge Network
← EdgeNet( , )
![Page 18: HEP.TrkX Charged Particle Tracking using Graph Neural Networks · HEP.TrkX Charged Particle Tracking using Graph Neural Networks. ... Long Short Term Memory - LSTM Breakthrough in](https://reader034.fdocuments.us/reader034/viewer/2022042220/5ec67716e690c36f2e14497e/html5/thumbnails/18.jpg)
03/10/19 18J.-R. Vlimant
Node Network
← NodeNet( , + , + + )
self incoming outgoing
![Page 19: HEP.TrkX Charged Particle Tracking using Graph Neural Networks · HEP.TrkX Charged Particle Tracking using Graph Neural Networks. ... Long Short Term Memory - LSTM Breakthrough in](https://reader034.fdocuments.us/reader034/viewer/2022042220/5ec67716e690c36f2e14497e/html5/thumbnails/19.jpg)
03/10/19 19J.-R. Vlimant
Information Flow● Graph is sparsely connected from layer to layer● InputNet + EdgeNet + NodeNet only correlates hits
information on triplet of layers✗ The information from the outer hits and inner hits
are not combined
● Several possible ways to operate the connection➔ Correlates hits information through multiple iterations of
(EdgeNet+NodeNet)
![Page 20: HEP.TrkX Charged Particle Tracking using Graph Neural Networks · HEP.TrkX Charged Particle Tracking using Graph Neural Networks. ... Long Short Term Memory - LSTM Breakthrough in](https://reader034.fdocuments.us/reader034/viewer/2022042220/5ec67716e690c36f2e14497e/html5/thumbnails/20.jpg)
03/10/19 20J.-R. Vlimant
Problem Size Considerations
![Page 21: HEP.TrkX Charged Particle Tracking using Graph Neural Networks · HEP.TrkX Charged Particle Tracking using Graph Neural Networks. ... Long Short Term Memory - LSTM Breakthrough in](https://reader034.fdocuments.us/reader034/viewer/2022042220/5ec67716e690c36f2e14497e/html5/thumbnails/21.jpg)
03/10/19 21J.-R. Vlimant
Dealing with Large Graphs
✗ Full event embedding✗ A graph with ~120k nodes (14.4B edges) and ~1M
potential edges is a big graph
● Split the problem➢ currently using 16 sectors in φ
● Use sparse matrix implementation ➢ https://github.com/deepmind/graph_nets for
example● Identify disjoint sub-graphs
➢ Geometrical cuts, segment pre-classifier, ...● Implement distributed learning of large graphs
➢ Scope of the Exa.TrkX Project
![Page 22: HEP.TrkX Charged Particle Tracking using Graph Neural Networks · HEP.TrkX Charged Particle Tracking using Graph Neural Networks. ... Long Short Term Memory - LSTM Breakthrough in](https://reader034.fdocuments.us/reader034/viewer/2022042220/5ec67716e690c36f2e14497e/html5/thumbnails/22.jpg)
03/10/19 22J.-R. Vlimant
Downgraded Complexity● TrackML dataset generated from ... with an
average of 200 pileup events.✗ Not computational possible at this time to
embed the smallest relevant sector of fullevent on a graph
➔ Sub-dataset are constructed by➢ Low density
✔ pT>1 GeV, Δφ<0.001, Δz
0<200mm
✔ acceptance: 99%, purity: 33%➢ Medium density
✔ pT>500 MeV, Δφ<0.0006, Δz
0<150mm
✔ acceptance: 95%, purity: 25%➢ High density
✔ pT>100 MeV, Δφ<0.0006, Δz
0<100mm
➔ acceptance: 43%, purity: 9%
![Page 23: HEP.TrkX Charged Particle Tracking using Graph Neural Networks · HEP.TrkX Charged Particle Tracking using Graph Neural Networks. ... Long Short Term Memory - LSTM Breakthrough in](https://reader034.fdocuments.us/reader034/viewer/2022042220/5ec67716e690c36f2e14497e/html5/thumbnails/23.jpg)
03/10/19 23J.-R. Vlimant
Performance
Low densityacc. x eff. ~ 97%
Medium densityacc x eff. ~ 90%
High densityacc. x eff. ~ 33%
![Page 24: HEP.TrkX Charged Particle Tracking using Graph Neural Networks · HEP.TrkX Charged Particle Tracking using Graph Neural Networks. ... Long Short Term Memory - LSTM Breakthrough in](https://reader034.fdocuments.us/reader034/viewer/2022042220/5ec67716e690c36f2e14497e/html5/thumbnails/24.jpg)
03/10/19 24J.-R. Vlimant
Summary
● Pilot project to explore new ideas forcharged particle track reconstruction
● Graph neural network show promisingresults even in increasingly dense event
● Post-processing, pre-processing, usingdomain knowledge, ... : work in progress
● Optimizing such models requires training atscale : issues to be tackled, stay tuned
![Page 25: HEP.TrkX Charged Particle Tracking using Graph Neural Networks · HEP.TrkX Charged Particle Tracking using Graph Neural Networks. ... Long Short Term Memory - LSTM Breakthrough in](https://reader034.fdocuments.us/reader034/viewer/2022042220/5ec67716e690c36f2e14497e/html5/thumbnails/25.jpg)
03/10/19 25J.-R. Vlimant
Part of this work was conducted at "iBanks", the AI GPU clusterat Caltech. We acknowledge NVIDIA, SuperMicro and theKavli Foundation for their support of "iBanks".
Acknowledgments
![Page 26: HEP.TrkX Charged Particle Tracking using Graph Neural Networks · HEP.TrkX Charged Particle Tracking using Graph Neural Networks. ... Long Short Term Memory - LSTM Breakthrough in](https://reader034.fdocuments.us/reader034/viewer/2022042220/5ec67716e690c36f2e14497e/html5/thumbnails/26.jpg)
03/10/19 26J.-R. Vlimant
Extra material
![Page 27: HEP.TrkX Charged Particle Tracking using Graph Neural Networks · HEP.TrkX Charged Particle Tracking using Graph Neural Networks. ... Long Short Term Memory - LSTM Breakthrough in](https://reader034.fdocuments.us/reader034/viewer/2022042220/5ec67716e690c36f2e14497e/html5/thumbnails/27.jpg)
03/10/19 27J.-R. Vlimant
Recall & Precision
Precision ≡ EfficiencyRecall ≡ Purity ≡ 1-(Fake rate)
Accuracy ≡ How much do we get it right
![Page 28: HEP.TrkX Charged Particle Tracking using Graph Neural Networks · HEP.TrkX Charged Particle Tracking using Graph Neural Networks. ... Long Short Term Memory - LSTM Breakthrough in](https://reader034.fdocuments.us/reader034/viewer/2022042220/5ec67716e690c36f2e14497e/html5/thumbnails/28.jpg)
03/10/19 28J.-R. Vlimant
Pattern Recognition With Deep Learning
![Page 29: HEP.TrkX Charged Particle Tracking using Graph Neural Networks · HEP.TrkX Charged Particle Tracking using Graph Neural Networks. ... Long Short Term Memory - LSTM Breakthrough in](https://reader034.fdocuments.us/reader034/viewer/2022042220/5ec67716e690c36f2e14497e/html5/thumbnails/29.jpg)
03/10/19 29J.-R. Vlimant
Machine Learning for Tracking
Zagoruyko et al, 1604.02135 Photo by Pier Marco Tacca/Getty Images
Many possible ways to cast the algorithm oftracking, or part of the current algorithms in a
machine learning problem
![Page 30: HEP.TrkX Charged Particle Tracking using Graph Neural Networks · HEP.TrkX Charged Particle Tracking using Graph Neural Networks. ... Long Short Term Memory - LSTM Breakthrough in](https://reader034.fdocuments.us/reader034/viewer/2022042220/5ec67716e690c36f2e14497e/html5/thumbnails/30.jpg)
03/10/19 30J.-R. Vlimant
Similarities and Challenges● Particle tracking is an active field in data science
➢ Different type of particles➢ Not oriented to code performance
● Making a track is a pattern recognition problem ➢ Not the usual one in data science
● Tracking data is much sparser than regular images➢ Test and adapt methods
● Tracking device may have up to 10M of channels➢ Scale up deep learning models➢ Perform tracking by sector
● Underlying geometry of sensor more complex➢ More than a simple picture➢ Barrels and end-caps are not the usual pictures
● Not the regular type of sequences➢ Cover new ground of sequence processing
● Defining an adequate cost function➢ Tracking algorithms are optimized by proxy
● A solution must be performant during inference ...
![Page 31: HEP.TrkX Charged Particle Tracking using Graph Neural Networks · HEP.TrkX Charged Particle Tracking using Graph Neural Networks. ... Long Short Term Memory - LSTM Breakthrough in](https://reader034.fdocuments.us/reader034/viewer/2022042220/5ec67716e690c36f2e14497e/html5/thumbnails/31.jpg)
03/10/19 31J.-R. Vlimant
HL-LHC Challenge
<PU>=7 <PU>=21
<PU>=140-20010x more hits
Circa 2025
● CPU time extrapolation into HL-LHC era far surpasses growth incomputing budget
● Need for faster algorithms
● Approximation allowed in the trigger
![Page 32: HEP.TrkX Charged Particle Tracking using Graph Neural Networks · HEP.TrkX Charged Particle Tracking using Graph Neural Networks. ... Long Short Term Memory - LSTM Breakthrough in](https://reader034.fdocuments.us/reader034/viewer/2022042220/5ec67716e690c36f2e14497e/html5/thumbnails/32.jpg)
03/10/19 32J.-R. Vlimant
Scene Labeling
From talk of LeCunn at CERN
![Page 33: HEP.TrkX Charged Particle Tracking using Graph Neural Networks · HEP.TrkX Charged Particle Tracking using Graph Neural Networks. ... Long Short Term Memory - LSTM Breakthrough in](https://reader034.fdocuments.us/reader034/viewer/2022042220/5ec67716e690c36f2e14497e/html5/thumbnails/33.jpg)
03/10/19 33J.-R. Vlimant
Scene Labeling
Farabet et al. ICML 2012, PAMI 2013
➔ Assign hits to track candidates
![Page 34: HEP.TrkX Charged Particle Tracking using Graph Neural Networks · HEP.TrkX Charged Particle Tracking using Graph Neural Networks. ... Long Short Term Memory - LSTM Breakthrough in](https://reader034.fdocuments.us/reader034/viewer/2022042220/5ec67716e690c36f2e14497e/html5/thumbnails/34.jpg)
03/10/19 34J.-R. Vlimant
Scene Captioning
Karpathy, Fei-Fei, CVPR 2015
➔ Compose tracks explanation from image
![Page 35: HEP.TrkX Charged Particle Tracking using Graph Neural Networks · HEP.TrkX Charged Particle Tracking using Graph Neural Networks. ... Long Short Term Memory - LSTM Breakthrough in](https://reader034.fdocuments.us/reader034/viewer/2022042220/5ec67716e690c36f2e14497e/html5/thumbnails/35.jpg)
03/10/19 35J.-R. Vlimant
Text Translation
➔ From sequence of hits on layer to sequence of hits on track
![Page 36: HEP.TrkX Charged Particle Tracking using Graph Neural Networks · HEP.TrkX Charged Particle Tracking using Graph Neural Networks. ... Long Short Term Memory - LSTM Breakthrough in](https://reader034.fdocuments.us/reader034/viewer/2022042220/5ec67716e690c36f2e14497e/html5/thumbnails/36.jpg)
03/10/19 36J.-R. Vlimant
Possible Application to Tracking● Track candidate
➔ Finding the hits that belong to a track➔ Seed + hits → tracks
● Track parameters➔ Measuring the physic quantity of tracks➔ Hits → track kinematics
● Seeding➔ Putting together hits into tracks➔ Hits → track
![Page 37: HEP.TrkX Charged Particle Tracking using Graph Neural Networks · HEP.TrkX Charged Particle Tracking using Graph Neural Networks. ... Long Short Term Memory - LSTM Breakthrough in](https://reader034.fdocuments.us/reader034/viewer/2022042220/5ec67716e690c36f2e14497e/html5/thumbnails/37.jpg)
03/10/19 37J.-R. Vlimant
HEP.TrkX Approaches
Seq-to-seq track finding
https://tinyurl.com/y87saehf
https://tinyurl.com/yb3v93y9
End-to-end hit assignment
Track following with RNN https://heptrkx.github.io/
![Page 38: HEP.TrkX Charged Particle Tracking using Graph Neural Networks · HEP.TrkX Charged Particle Tracking using Graph Neural Networks. ... Long Short Term Memory - LSTM Breakthrough in](https://reader034.fdocuments.us/reader034/viewer/2022042220/5ec67716e690c36f2e14497e/html5/thumbnails/38.jpg)
03/10/19 38J.-R. Vlimant
Seeded Track Candidate Making
![Page 39: HEP.TrkX Charged Particle Tracking using Graph Neural Networks · HEP.TrkX Charged Particle Tracking using Graph Neural Networks. ... Long Short Term Memory - LSTM Breakthrough in](https://reader034.fdocuments.us/reader034/viewer/2022042220/5ec67716e690c36f2e14497e/html5/thumbnails/39.jpg)
03/10/19 39J.-R. Vlimant
Seeded Pattern Prediction
● Hits on first 3 layers are used as seed● Predict the position of the rest of the hits on all layers
![Page 40: HEP.TrkX Charged Particle Tracking using Graph Neural Networks · HEP.TrkX Charged Particle Tracking using Graph Neural Networks. ... Long Short Term Memory - LSTM Breakthrough in](https://reader034.fdocuments.us/reader034/viewer/2022042220/5ec67716e690c36f2e14497e/html5/thumbnails/40.jpg)
03/10/19 40J.-R. Vlimant
Long Short Term Memory - LSTM
Breakthrough in sequence processing by carrying overan internal state, “memory” of the previous items in the
sequence, allowing for long range correlation
http://colah.github.io/posts/2015-08-Understanding-LSTMs/
![Page 41: HEP.TrkX Charged Particle Tracking using Graph Neural Networks · HEP.TrkX Charged Particle Tracking using Graph Neural Networks. ... Long Short Term Memory - LSTM Breakthrough in](https://reader034.fdocuments.us/reader034/viewer/2022042220/5ec67716e690c36f2e14497e/html5/thumbnails/41.jpg)
03/10/19 41J.-R. Vlimant
LSTM ≡ Kalman Filter
![Page 42: HEP.TrkX Charged Particle Tracking using Graph Neural Networks · HEP.TrkX Charged Particle Tracking using Graph Neural Networks. ... Long Short Term Memory - LSTM Breakthrough in](https://reader034.fdocuments.us/reader034/viewer/2022042220/5ec67716e690c36f2e14497e/html5/thumbnails/42.jpg)
03/10/19 42J.-R. Vlimant
Seeded Pattern Recognition Insights● For a simplified track models,
predicting the track pattern from theseed works
➢ In 2D and 3D➢ With some level of noise➢ With other tracks present➢ On layers with increasing number
of pixels
● Several other architectures tried➢ Convolutional neural nets
(no LSTM)➢ Convolutional auto-encoder➢ Bi-directional LSTM➢ Prediction on next layer with LSTM
![Page 43: HEP.TrkX Charged Particle Tracking using Graph Neural Networks · HEP.TrkX Charged Particle Tracking using Graph Neural Networks. ... Long Short Term Memory - LSTM Breakthrough in](https://reader034.fdocuments.us/reader034/viewer/2022042220/5ec67716e690c36f2e14497e/html5/thumbnails/43.jpg)
03/10/19 43J.-R. Vlimant
Tracking RAMP at CtDS. Farrell : Best solution in the Machine Learning categoryhttps://indico.cern.ch/event/577003/contributions/2509988/
● Down-sampling layer to 100 bins● LSTM for hit assignment● 92% efficiency● Robust to holes and missing hits
● Increased granularity in “road”● LSTM for hit assignment● 95% efficiency
![Page 44: HEP.TrkX Charged Particle Tracking using Graph Neural Networks · HEP.TrkX Charged Particle Tracking using Graph Neural Networks. ... Long Short Term Memory - LSTM Breakthrough in](https://reader034.fdocuments.us/reader034/viewer/2022042220/5ec67716e690c36f2e14497e/html5/thumbnails/44.jpg)
03/10/19 44J.-R. Vlimant
Finding Tracks with LSTMLSTM ≡ Kalman Filter
➢ Search seeded from aknown tracklet
➢ Hit location is discretized tofixed length
➢ Model predicts the binnedposition of the hit on thenext layer
![Page 45: HEP.TrkX Charged Particle Tracking using Graph Neural Networks · HEP.TrkX Charged Particle Tracking using Graph Neural Networks. ... Long Short Term Memory - LSTM Breakthrough in](https://reader034.fdocuments.us/reader034/viewer/2022042220/5ec67716e690c36f2e14497e/html5/thumbnails/45.jpg)
03/10/19 45J.-R. Vlimant
Hit Prediction with Gaussian Model
➢ Search seeded from aknown tracklet
➢ Hit positions taken insequential input
➢ Model predicts theposition of the hit on thenext layer
Loss function incorporates the position andthe predicted uncertainty
![Page 46: HEP.TrkX Charged Particle Tracking using Graph Neural Networks · HEP.TrkX Charged Particle Tracking using Graph Neural Networks. ... Long Short Term Memory - LSTM Breakthrough in](https://reader034.fdocuments.us/reader034/viewer/2022042220/5ec67716e690c36f2e14497e/html5/thumbnails/46.jpg)
03/10/19 46J.-R. Vlimant
Track Parameters Measurement
![Page 47: HEP.TrkX Charged Particle Tracking using Graph Neural Networks · HEP.TrkX Charged Particle Tracking using Graph Neural Networks. ... Long Short Term Memory - LSTM Breakthrough in](https://reader034.fdocuments.us/reader034/viewer/2022042220/5ec67716e690c36f2e14497e/html5/thumbnails/47.jpg)
03/10/19 47J.-R. Vlimant
Track Parameter Estimation
Try to predict theslope and intersectof this track
![Page 48: HEP.TrkX Charged Particle Tracking using Graph Neural Networks · HEP.TrkX Charged Particle Tracking using Graph Neural Networks. ... Long Short Term Memory - LSTM Breakthrough in](https://reader034.fdocuments.us/reader034/viewer/2022042220/5ec67716e690c36f2e14497e/html5/thumbnails/48.jpg)
03/10/19 48J.-R. Vlimant
Multi-Track Prediction with LSTM
● Hit pattern from multiple trackprocessed through convolutionallayers
● LSTM Cell runs for as manytracks the model can predict.
![Page 49: HEP.TrkX Charged Particle Tracking using Graph Neural Networks · HEP.TrkX Charged Particle Tracking using Graph Neural Networks. ... Long Short Term Memory - LSTM Breakthrough in](https://reader034.fdocuments.us/reader034/viewer/2022042220/5ec67716e690c36f2e14497e/html5/thumbnails/49.jpg)
03/10/19 49J.-R. Vlimant
Predicting Covariance Matrix
● The observed hit pattern from multiple trackprocessed through convolutional layers
● LSTM cells are ran multiple time in order to predicta list of particles
● Model is able to predict the covariance matrix oftrack parameters, incorporated in the loss function
![Page 50: HEP.TrkX Charged Particle Tracking using Graph Neural Networks · HEP.TrkX Charged Particle Tracking using Graph Neural Networks. ... Long Short Term Memory - LSTM Breakthrough in](https://reader034.fdocuments.us/reader034/viewer/2022042220/5ec67716e690c36f2e14497e/html5/thumbnails/50.jpg)
03/10/19 50J.-R. Vlimant
Track Parameter Prediction
Graphical representation of track slope,intersect and respective uncertainties
Hit pattern inthe detector
Track parameters andcorrespondinguncertainties
![Page 51: HEP.TrkX Charged Particle Tracking using Graph Neural Networks · HEP.TrkX Charged Particle Tracking using Graph Neural Networks. ... Long Short Term Memory - LSTM Breakthrough in](https://reader034.fdocuments.us/reader034/viewer/2022042220/5ec67716e690c36f2e14497e/html5/thumbnails/51.jpg)
03/10/19 51J.-R. Vlimant
Hit Assignment Approaches
![Page 52: HEP.TrkX Charged Particle Tracking using Graph Neural Networks · HEP.TrkX Charged Particle Tracking using Graph Neural Networks. ... Long Short Term Memory - LSTM Breakthrough in](https://reader034.fdocuments.us/reader034/viewer/2022042220/5ec67716e690c36f2e14497e/html5/thumbnails/52.jpg)
03/10/19 52J.-R. Vlimant
Pattern Recognition
Try to assemblehits into trackcandidates.
![Page 53: HEP.TrkX Charged Particle Tracking using Graph Neural Networks · HEP.TrkX Charged Particle Tracking using Graph Neural Networks. ... Long Short Term Memory - LSTM Breakthrough in](https://reader034.fdocuments.us/reader034/viewer/2022042220/5ec67716e690c36f2e14497e/html5/thumbnails/53.jpg)
03/10/19 53J.-R. Vlimant
seq-2-seq tracking● Input sequence of hits per layers (one sequence per layer)
➢ One LSTM cell per layer● Output sequence of hits per candidates
➢ Final LSTM runs for as many candidates the model can predict
Restricted to 4 layers(with seeding in mind)
Full performanceevaluation still to be done
![Page 54: HEP.TrkX Charged Particle Tracking using Graph Neural Networks · HEP.TrkX Charged Particle Tracking using Graph Neural Networks. ... Long Short Term Memory - LSTM Breakthrough in](https://reader034.fdocuments.us/reader034/viewer/2022042220/5ec67716e690c36f2e14497e/html5/thumbnails/54.jpg)
03/10/19 54J.-R. Vlimant
Hit Assignment Algorithm
➢ Unseeded hit-to-track assignment (clustering)➢ Hit positions taken in sequential input➢ Model predicts the probability that a hit belongs
to a track candidate
![Page 55: HEP.TrkX Charged Particle Tracking using Graph Neural Networks · HEP.TrkX Charged Particle Tracking using Graph Neural Networks. ... Long Short Term Memory - LSTM Breakthrough in](https://reader034.fdocuments.us/reader034/viewer/2022042220/5ec67716e690c36f2e14497e/html5/thumbnails/55.jpg)
03/10/19 55J.-R. Vlimant
Vertexing
![Page 56: HEP.TrkX Charged Particle Tracking using Graph Neural Networks · HEP.TrkX Charged Particle Tracking using Graph Neural Networks. ... Long Short Term Memory - LSTM Breakthrough in](https://reader034.fdocuments.us/reader034/viewer/2022042220/5ec67716e690c36f2e14497e/html5/thumbnails/56.jpg)
03/10/19 56J.-R. Vlimant
Vertexing with CNN
➢ Using hits binned (η, φ) map ininput for a regression of theprimary vertex position
➢ Modest success
![Page 57: HEP.TrkX Charged Particle Tracking using Graph Neural Networks · HEP.TrkX Charged Particle Tracking using Graph Neural Networks. ... Long Short Term Memory - LSTM Breakthrough in](https://reader034.fdocuments.us/reader034/viewer/2022042220/5ec67716e690c36f2e14497e/html5/thumbnails/57.jpg)
03/10/19 57J.-R. Vlimant
Graph Networks Approach
![Page 58: HEP.TrkX Charged Particle Tracking using Graph Neural Networks · HEP.TrkX Charged Particle Tracking using Graph Neural Networks. ... Long Short Term Memory - LSTM Breakthrough in](https://reader034.fdocuments.us/reader034/viewer/2022042220/5ec67716e690c36f2e14497e/html5/thumbnails/58.jpg)
03/10/19 58J.-R. Vlimant
Seeded Hit Classification with GNN➢ Seeded hit
classification➢ Model predicts
whether hits belongto the given seed
![Page 59: HEP.TrkX Charged Particle Tracking using Graph Neural Networks · HEP.TrkX Charged Particle Tracking using Graph Neural Networks. ... Long Short Term Memory - LSTM Breakthrough in](https://reader034.fdocuments.us/reader034/viewer/2022042220/5ec67716e690c36f2e14497e/html5/thumbnails/59.jpg)
03/10/19 59J.-R. Vlimant
Track Building With GNN
See our poster on Track 6 for more detailshttps://indico.cern.ch/event/587955/contributions/2937570/
Successive iterations on a selected event
➢ Unseeded hit-pair classification➢ Model predicts the probability that a hit-pair is valid
![Page 60: HEP.TrkX Charged Particle Tracking using Graph Neural Networks · HEP.TrkX Charged Particle Tracking using Graph Neural Networks. ... Long Short Term Memory - LSTM Breakthrough in](https://reader034.fdocuments.us/reader034/viewer/2022042220/5ec67716e690c36f2e14497e/html5/thumbnails/60.jpg)
03/10/19 60J.-R. Vlimant
Hardware Consideration
![Page 61: HEP.TrkX Charged Particle Tracking using Graph Neural Networks · HEP.TrkX Charged Particle Tracking using Graph Neural Networks. ... Long Short Term Memory - LSTM Breakthrough in](https://reader034.fdocuments.us/reader034/viewer/2022042220/5ec67716e690c36f2e14497e/html5/thumbnails/61.jpg)
03/10/19 61J.-R. Vlimant
Inference on FPGA● Demo at NIPS 2017 of implementing neural
networks on FPGA● Collaborating with hls4ml team to push the
graph neural networks models to the nexts level
See Jennifer's talk during this event https://indico.cern.ch/event/587955/contributions/2937529/
![Page 62: HEP.TrkX Charged Particle Tracking using Graph Neural Networks · HEP.TrkX Charged Particle Tracking using Graph Neural Networks. ... Long Short Term Memory - LSTM Breakthrough in](https://reader034.fdocuments.us/reader034/viewer/2022042220/5ec67716e690c36f2e14497e/html5/thumbnails/62.jpg)
03/10/19 62J.-R. Vlimant
Tracking Not In a Nutshell
● Hits preparation
● Seeding
● Pattern recognition
● Track fitting
● Track cleaning
Sev
eral
Tim
es
![Page 63: HEP.TrkX Charged Particle Tracking using Graph Neural Networks · HEP.TrkX Charged Particle Tracking using Graph Neural Networks. ... Long Short Term Memory - LSTM Breakthrough in](https://reader034.fdocuments.us/reader034/viewer/2022042220/5ec67716e690c36f2e14497e/html5/thumbnails/63.jpg)
03/10/19 63J.-R. Vlimant
Hit Preparation
● Calculate the hit position from barycenter of chargedeposits
● Use of neural net classifier to split cluster in ATLAS
● Access to trajectory local parameter from clustershape
● Remove hits from previous tracking iterations
● HL-LHC design include double layers giving moreconstraints on the local trajectory parameters
![Page 64: HEP.TrkX Charged Particle Tracking using Graph Neural Networks · HEP.TrkX Charged Particle Tracking using Graph Neural Networks. ... Long Short Term Memory - LSTM Breakthrough in](https://reader034.fdocuments.us/reader034/viewer/2022042220/5ec67716e690c36f2e14497e/html5/thumbnails/64.jpg)
03/10/19 64J.-R. Vlimant
Seeding
● Combinatorics of 2 or 3 hitswith tight/loose constraintsto the beam spot or vertex
● Seed cleaning/purity playsin an important in reducingthe CPU requirements ofsub-sequent steps➔ Consider pixel cluster
shape and charge toremove incompatibleseeds
● Initial track parameters fromhelix fit
![Page 65: HEP.TrkX Charged Particle Tracking using Graph Neural Networks · HEP.TrkX Charged Particle Tracking using Graph Neural Networks. ... Long Short Term Memory - LSTM Breakthrough in](https://reader034.fdocuments.us/reader034/viewer/2022042220/5ec67716e690c36f2e14497e/html5/thumbnails/65.jpg)
03/10/19 65J.-R. Vlimant
Pattern Recognition● Use of the Kalman filter
formalism with weight matrix
● Identify possible next layersfrom geometrical considerations
● Combinatorics with compatibleshits, retain N best candidates
● No smoothing procedure
● Resilient to missing modules
● Hits are mostly belonging to onetrack and one track only
● Hit sharing can happen in denseevents, in the innermost part
● Lots of hits from low momentumparticles
![Page 66: HEP.TrkX Charged Particle Tracking using Graph Neural Networks · HEP.TrkX Charged Particle Tracking using Graph Neural Networks. ... Long Short Term Memory - LSTM Breakthrough in](https://reader034.fdocuments.us/reader034/viewer/2022042220/5ec67716e690c36f2e14497e/html5/thumbnails/66.jpg)
03/10/19 66J.-R. Vlimant
Kalman Filter● Trajectory state propagation
done either✔ Analytical (helix, fastest)✔ Stepping helix (fast)✔ Runge-Kutta (slow)
● Material effect added totrajectory state covariance
● Projection matrix of local helixparameters onto module surface➔ Trivial expression due to local
helix parametrisation● Hits covariance matrix for pixel
and stereo hits properly formed✗ Issue with strip hits and
longitudinal error being nongaussian (square)
![Page 67: HEP.TrkX Charged Particle Tracking using Graph Neural Networks · HEP.TrkX Charged Particle Tracking using Graph Neural Networks. ... Long Short Term Memory - LSTM Breakthrough in](https://reader034.fdocuments.us/reader034/viewer/2022042220/5ec67716e690c36f2e14497e/html5/thumbnails/67.jpg)
03/10/19 67J.-R. Vlimant
Track Fitting
● Use of the Kalman filterformalism with weightmatrix
● Use of smoothingprocedure to identifyoutliers
● Field non uniformity aretaken into account
● Detector alignmenttaken into account
![Page 68: HEP.TrkX Charged Particle Tracking using Graph Neural Networks · HEP.TrkX Charged Particle Tracking using Graph Neural Networks. ... Long Short Term Memory - LSTM Breakthrough in](https://reader034.fdocuments.us/reader034/viewer/2022042220/5ec67716e690c36f2e14497e/html5/thumbnails/68.jpg)
03/10/19 68J.-R. Vlimant
Cleaning, Selection
● Track qualityestimated usingranking orclassification method➔Use of MVA
● Hits from high qualitytracks are removefor the next iterationswhere applicable
![Page 69: HEP.TrkX Charged Particle Tracking using Graph Neural Networks · HEP.TrkX Charged Particle Tracking using Graph Neural Networks. ... Long Short Term Memory - LSTM Breakthrough in](https://reader034.fdocuments.us/reader034/viewer/2022042220/5ec67716e690c36f2e14497e/html5/thumbnails/69.jpg)
03/10/19 69J.-R. Vlimant
A Charged Particle Journey
![Page 70: HEP.TrkX Charged Particle Tracking using Graph Neural Networks · HEP.TrkX Charged Particle Tracking using Graph Neural Networks. ... Long Short Term Memory - LSTM Breakthrough in](https://reader034.fdocuments.us/reader034/viewer/2022042220/5ec67716e690c36f2e14497e/html5/thumbnails/70.jpg)
03/10/19 70J.-R. Vlimant
First order effect : electromagnetic elasticinteraction of the charge particle with nuclei (heavy
and multiply charged) and electrons (light andsingle charged)
Second order effect : inelastic interaction withnuclei.
![Page 71: HEP.TrkX Charged Particle Tracking using Graph Neural Networks · HEP.TrkX Charged Particle Tracking using Graph Neural Networks. ... Long Short Term Memory - LSTM Breakthrough in](https://reader034.fdocuments.us/reader034/viewer/2022042220/5ec67716e690c36f2e14497e/html5/thumbnails/71.jpg)
03/10/19 71J.-R. Vlimant
Magnetic Field● Magnetic fieldB acts on charged particles
in motion : Lorentz Force
● The solution in uniform magnetic field isan helix along the field : 5 parameters
● Helix radius proportional to the componentof momentum perpendicular to B
● Separate particles in dense environment
➔ Bending induces radiation :bremsstrahlung
➔ The magnetic field has to be known to agood precision for accurate tracking ofparticle
![Page 72: HEP.TrkX Charged Particle Tracking using Graph Neural Networks · HEP.TrkX Charged Particle Tracking using Graph Neural Networks. ... Long Short Term Memory - LSTM Breakthrough in](https://reader034.fdocuments.us/reader034/viewer/2022042220/5ec67716e690c36f2e14497e/html5/thumbnails/72.jpg)
03/10/19 72J.-R. Vlimant
Multiple Scattering● Deflection on nuclei (effect from
electron are negligible)
● Addition of scattering processes
● Gaussian approximation valid forsubstantial material traversed
Gaussian Approximation
![Page 73: HEP.TrkX Charged Particle Tracking using Graph Neural Networks · HEP.TrkX Charged Particle Tracking using Graph Neural Networks. ... Long Short Term Memory - LSTM Breakthrough in](https://reader034.fdocuments.us/reader034/viewer/2022042220/5ec67716e690c36f2e14497e/html5/thumbnails/73.jpg)
03/10/19 73J.-R. Vlimant
Bremsstrahlung● Electromagnetic radiation of
charged particles under accelerationdue to nuclei charge
● Significant at low mass or highenergy
● Discontinuity in energy lossspectrum due to photon emissionand track curvature
➔ Can be observed as kink in thetrajectory or presence of collinearenergetic photons
![Page 74: HEP.TrkX Charged Particle Tracking using Graph Neural Networks · HEP.TrkX Charged Particle Tracking using Graph Neural Networks. ... Long Short Term Memory - LSTM Breakthrough in](https://reader034.fdocuments.us/reader034/viewer/2022042220/5ec67716e690c36f2e14497e/html5/thumbnails/74.jpg)
03/10/19 74J.-R. Vlimant
Energy Loss● Momentum transfer to electrons when
traversing material (effect of nuclei isnegligible
● Energy loss at low momentumdepends on mass : can be used asmass spectrometer
ALICE Experiment
![Page 75: HEP.TrkX Charged Particle Tracking using Graph Neural Networks · HEP.TrkX Charged Particle Tracking using Graph Neural Networks. ... Long Short Term Memory - LSTM Breakthrough in](https://reader034.fdocuments.us/reader034/viewer/2022042220/5ec67716e690c36f2e14497e/html5/thumbnails/75.jpg)
03/10/19 75J.-R. Vlimant
Summary on Material Effects
● Collective effects can be estimatedstatistically and taken into account in how theymodify the trajectory
● Bremstrahlung and nuclear interactionssignificantly distort trajectories