Machine Learning in High Energy Physics | 9am PT Feb 19

[Grigory Bronevetsky]

Modeling Talks

Machine Learning in High Energy Physics V Hewes, U Cincinnati Tuesday, Feb 19 | 9am PT Meet | Youtube Stream Hi all, The presentation will be via Meet and all questions will be addressed there. If you cannot attend live, the event will be recorded and can be found afterward at https://sites.google.com/modelingtalks.org/entry/machine-learning-in-high-energy-physics 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. More information on previous and future talks: https://sites.google.com/modelingtalks.org/entry/home

[Grigory Bronevetsky]

Video Recording: https://youtu.be/pVOc_qNtS5M

Slides: https://drive.google.com/file/d/1HrqQ-r85HvMdmtp9MXTiTThgpJ2kjRZ6/view?usp=sharing

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