Skip to content

Latest commit

 

History

History
36 lines (32 loc) · 2.78 KB

Usecases-of-GNN.md

File metadata and controls

36 lines (32 loc) · 2.78 KB

Particle Track Reconstruction in Large Hadron Collider (LHC)

Farrell, Steven, et al. "Novel deep learning methods for track reconstruction." arXiv preprint arXiv:1810.06111 (2018).

In LHC, data are produced at an unprecendented scale. One of the important tasks is to track particle as they move around the collider. GNN is being used for particle tracking.

  • Data are space-point representation of HEPtrackingdata
  • Contains Hits, interactions information
  • Track building with Recurrent Neural Networks(RNNs)
  • Particle tracks can be represented as a sequence of hits
  • Given a sequence of hit coordinates, the model produces for every element a prediction of the position of the next hit conditioned on its position and the preceding hit positions.
    • A track seed using the first three hits of a true track and the RNN models are used to make remaining predictions
    • It selects the highest scoring hit in the event on each successive layer
  • Track finding with Graph Neural Networks (GNNs)

    • How graph is constructed particle-graph

      • Connects hits using geo-metric constraints or by using preprocessing algorithm
      • Hits are also connected on adjacent layers when they are compatible according to some criteria
      • During inference, graphs are constructed with 4 hits on each detector layer in the region around the true track and connecting all hits together on adjacent layers.

    • Architecture of GNN model

      • An input transformation layer appears first and then is followed by alternating EdgeNetwork and NodeNetwork
      • NodeNetwork computes node features based on the neighboring nodes and weight neighboring nodes information using edge weight.
      • EdgeNetwork computes edge's weight based on the start and end node
      • In each iteration of the GNN, the model propagates information through the graph that helps to stress important connections and weaken useless connections.

    • Identifies a track by performing binary hit classification in a partially labeled graph. It is node classification.

      • Graph is given first three hit labels as True. 7 iteration of information passing is performed. Sigmoid activation is done to identify node label (True or False) denoting whether node belongs to the track.

    • Identifies many tracks at once by performing binary segment classification. It performs classification on graph edges

    • Scaling

      • Subset of datasets are created maining a balance between small, medium and large training files
      • Distributed training have been applied.

Reference:

  1. Poster
  2. Farrell, Steven, et al. "Novel deep learning methods for track reconstruction." arXiv preprint arXiv:1810.06111 (2018).