Abstract:
Machine Learning (ML) algorithms are increasingly being adopted to address a broad variety of challenges in experimental High Energy Physics (HEP), including event classification, clustering, particle identification and more. This talk charts the evolution of ML in HEP, from the deployment of standard image recognition architectures to the development of advanced graph neural network (GNN) and sparse convolution neural network (SCNN) applications. Particular reference is taken to two major particle physics experiments, the Deep Underground Neutrino Experiment (DUNE) and the High-Luminosity Large Hadron Collider (HL-LHC), in order to highlight both their commonalities and the unique challenges posed by the neutrino and collider paradigms. These approaches are then situated within the context of a typical particle physics analysis, which requires robust error estimation and quantification of an algorithm's inefficiency and bias.
Bio:
V Hewes is a neutrino physics researcher at the University of Cincinnati, based at the Fermi National Accelerator Laboratory (Fermilab). She is a member of the NOvA collaboration, on which she conducts neutrino oscillation analysis and acts as computing coordinator. She also collaborates on the next-generation DUNE experiment, for which she develops GNN reconstruction techniques for Liquid Argon detectors as part of the ExaTrkX collaboration. Her primary research interests lie at the intersection of experimental physics and scientific computing, developing generalized tools for reconstruction and analysis that can be leveraged across a diverse range of experiments and architectures.
Summary
Focus: Use of ML in High Energy physics (HEP)
Neutrino physics and experiments
3 Types: Electron, Mu, Tau Neutrionos
Know that there are exactly 3 from the Large Electron Positron Collider
Solar neutrino problem:
Homestake experiment: flux of neutrinos emitted by the sun
Found that there were far fewer than theory predicted
Resolved by experiments that showed that neutrino flavors mutate from electron to mu and tau on the way from Sun to Earth
New theory of how neutrinos oscillate, shows that neutrinos have mass because they must experience time
Neutrino oscillations are described via 6 parameters
Probability depends on the path length (how far its traveled) and energy of the neutrino
Intensity Frontier (many neutrinos): Deep Underground Neutrino Experiment (DUNE)
Shoots a beam of neutrinos underground from Fermilab to a detector in South Dakota (1 mile underground)
Detector: Liquid Argon Time Projection Chambers (LArTPCs)
Charged particles ionize liquid argon as they travel
Particle tracks detected at 3mm spatial resolution
Energy Frontier (high-energy neutrinos): Large Hadron Collider (LHC)
Proton-proton collisions at 14TeV
Analysis:
Data too complicated to analyze directly
Define test statistics that include both observed data and underlying physics parameter of interest
Neutrino physics: Poisson likelihood
Need simulation that
Predicts the outcome of the experiment for various values of the physical parameters
Very high accuracy, accounts for details of the detector, the beamline, the distribution of particles in the beam, etc.
Stages:
Event generation (GENIE: neutrino event generator: https://hep.ph.liv.ac.uk/~costasa/genie/)
Models the initial interaction of neutrinos with atomic nuclei to generate distribution of primary neutrinos
E.g. neutrino -> neutron -> proton + muon
Particle tracking (Geant4: https://geant4.web.cern.ch/)
Time steps primary to predict how they flow through materials over time
Predicts their ultimate fates as they arrive at and interact with the detector, and deposit energy within it
Detector simulation
Custom simulation for each experiment
For LArTPC:
Electron drift, recombination, etc.
Wire response, electronics, etc.
Prediction: raw waveforms on the detector wires
Reconstruction: go from observed data backwards through simulation chain to understand the phenomena that must have led to these observations
Construct raw hits from raw waveforms
Combine energy measurements on different 2D banks of wires
Cluster these measurements into events (usually geometric, rather than physical)
Infer the particle that must have caused each event
Aggregate across many events to compute/characterize flux
The many steps of this reconstruction loses accuracy bit by bit
A fully end-to-end supervised ML-based reconstruction can improve accuracy
Open datasets make this possible
LHC TrackML dataset: https://sites.google.com/site/trackmlparticle/home
MicroBooNE open data: https://microboone.fnal.gov/documents-publications/public-datasets/
Common ML Tasks
Event classification: observation -> particles (neutrino flavor, signal vs background)
Particle reconstruction: grouping detector hist, identification
Parameter estimation: directionality, vertexing, etc.
History of ML in HEP
CNN
Popular because many detectors represent 3D phenomena using 2D maps
NOvA uses CNNs to identify neutrino candidates
DUNE uses CNNs as selection for oscillation sensitivities
NOvA uses particle-level CNN to identify clusters within event
ProtoDONE: track-shower separation
Major limitations: data is spatially sparse, which is a waste for CNNs, which are structurally dense
Sparse CNNs explored: e.g. encoded/decoder architecture
Sparse voxel map has no need for truncation or downsampling
GNN
Makes it possible to organize the network around the structure of the problem
IceCub collaboration
Sheet of antarctic ice kms in size as detector
Drop detectors of high energy particles deep into ice in irregular structure
Nodes as discrete detector modules
HEP.TrkX: nodes as hits, predict links between hits in sequential detector layers
NuGraph: general purpose GNN for particle reconstruction for neutrinos https://github.com/exatrkx/NuGraph
NuGraph2:
Describe hit on each detector plane as heterogeneous graph
Using Delaunay triangulations to connect hits within each plane
Self-attention message passing network, first within plane, then across planes
Good performance due to good use of sparsity
Background filtering: ~97% accuracy
Hit classification: ~95% accuracy
NuGraph3: aim for full reconstruction of particle hierarchy/shower that is fully differentiable and complete
HL-LHC ExaTrkX pipeline: graph is shaped radially like the detector
Object condensation loss: tries to improve clustering by modeling attractive and repulsive potentials for event, can be used in DBScan
Particle flow reconstruction: full particle hierarchy / shower
GrapPA: sparse CNNs for full particle reconstruction