ProtoDUNE Performance and Analysis Utilities Leigh Whitehead DUNE - PowerPoint PPT Presentation

ProtoDUNE Performance and Analysis Utilities Leigh Whitehead DUNE UK Meeting 11/12/19

Overview • This talk will be in two main sections • Pandora performance in ProtoDUNE • A lot of this was Steve’s work • Analysis-level utilities for accessing the outputs of the reconstruction Leigh Whitehead 2

Pandora in ProtoDUNE • Pandora is the only particle reconstruction algorithm used in ProtoDUNE-SP • Generally, it has been working very well! • All of the future improvements that will be discussed today will improve the performance • I’ll show a few slides presented by Steve Green at the last DUNE collaboration meeting • Represents the current status as of the second data processing this summer Leigh Whitehead 3

Pandora in ProtoDUNE • Reconstructed data event in the upcoming ProtoDUNE-SP performance paper • Looks like a nice pion 
 charge exchange event • See beam particles amongst 
 large cosmic background w , wire position Beam Interaction Cosmic muons x , drift position Leigh Whitehead 4

Pandora in ProtoDUNE Things look very similar for protons Leigh Whitehead 5

Pandora in ProtoDUNE Leigh Whitehead 6

Pandora in ProtoDUNE Leigh Whitehead 7

Challenges for ProtoDUNE • Most of the analysis challenges are sub-leading effects • Identification of short tracks emerging from the beam particle interactions • Detection of small angle kinks for cross-section measurements (from elastic scattering) • Beam interaction vertex finding • The current activities underway for the DUNE FD will also have a strong impact on these issues Leigh Whitehead 8

Pandora for DUNE FD • We have to ensure that Pandora remains at the forefront of the DUNE FD event reconstruction • Very well represented in the DUNE TDR • Hence, we must: • Present updates regularly so that people know it exists and is under active development • Help people to use it, and its outputs, correctly and easily • We are in a good position to do the former: • The UK has two of the DUNE FD Sim/Reco convenors and the ProtoDUNE-SP reco convenor Leigh Whitehead 9

Analysis Utilities • The aim is to provide easy access to the reconstructed information • This also allows us to ensure the analysers access the information in our recommended way • Provide intuitive access to the particle hierarchy and break the “look at tracks and blobs” mentality we keep with us from our older experiments • I also try to hide some of the annoying things about LArSoft • I wrote similar utilities for ProtoDUNE and they are widely used by the analysers Leigh Whitehead 10


 Recommended Workflow Reconstructed objects should be Event accessed using the hierarchy Always start with the PFParticles PFParticle … Track Cluster Shower … … Hits … Recommended Following each arrow is a few Available but discouraged lines of code in LArSoft Leigh Whitehead 11


 Recommended Workflow Reconstructed objects should be Event accessed using the hierarchy Always start with the PFParticles PFParticle Access other products via the PFParticles … Track Cluster Shower … … Hits … Recommended Following each arrow is a few Available but discouraged lines of code in LArSoft Leigh Whitehead 12


 Recommended Workflow Provide simple functions to Event access these products in the recommended way PFParticle … Track Cluster Shower … … Hits … Recommended Following each arrow is a few Available but discouraged lines of code in LArSoft Leigh Whitehead 13


 Recommended Workflow I will provide functions to skip Event these steps for high level users, but provide warnings that it is not the standard method PFParticle … Track Cluster Shower … … Hits … Recommended Following each arrow is a few Available but discouraged lines of code in LArSoft Leigh Whitehead 14

The Utilities • I have written a number of these utilities • Here’s a quick look at the PFParticle one larpandora/LArPandoraAnalysisUtils/LArPandoraPFParticleUtils.h • Easy access to the products that make up the PFParticle Leigh Whitehead 15

The Utilities • There are utilities for accessing information from: • Events • PFParticles • Tracks • Showers • Clusters • Slices • SpacePoints • Code is easily maintainable since there are only three underlying functions that do the work • Extendable to any (current or future) LArSoft products via templated functions Leigh Whitehead 16

The Utilities • Allows for quick and simple code • Two lines to get the reconstructed neutrino and interaction products art::Ptr<recob::PFParticle> nu = LArPandoraEventUtils::GetNeutrino(event,”pandora”); std::vector<art::Ptr<recob::PFParticle>> children = LArPandoraPFParticleUtils::GetChildParticles(nu,event,”pandora”); • We can then get the interaction vertex art::Ptr<recob::Vertex> nuVertex = LArPandoraPFParticleUtils::GetVertex(nu,event,”pandora”); • Look for track-like child particles for(art::Ptr<recob::PFParticle> &child : children) 
 { if (LArPandoraPFParticleUtils::IsTrack(child,event,”pandora”,”pandoraTrack”)) { <insert user code here> } } Leigh Whitehead 17

Summary • Pandora is performing well for ProtoDUNE, but there is always room for improvement! • Moving forwards we have to ensure Pandora is the primary reconstruction tool • We all need to present update regularly • Analysis-level tools will ensure that users can access the information from Pandora and using it in the best way Leigh Whitehead 18

Bonus: Track PID • I take the final 100 hits of pandora tracks as 1D arrays • Extracted from LArSoft using the analysis utilities • If a track has 30 < n_hits < 100 then pad the start of the track as follows: • Calculate the mean and sigma of dE/dx for hits 10 to 20 • Pad the array to length 100 using random values: Gauss(mean,sigma) • Store some additional variables • Mean and sigma of dE/dx for hits 10 to 20 • Sigma of the distribution of angles between successive track points (how wobbly it is) • Number of track-like and shower-like child particles Leigh Whitehead 19

Deep Learning in 1D • There are a few standard techniques for processing 1D information… lots of examples from speech recognition etc • I’ve investigated the use of 1D convolutions so far 100 element dE/dx array 1D Convolutions 1D Convolutions 1D Convolutions 1D Convolutions Fully Connected Layer Fully Connected Layer Fully Connected Layer Fully Connected Layer Outputs # Track-like Children # Shower-like Children Mean dE/dx (10 - 20) Sigma dE/dx (10 - 20) Wobble Leigh Whitehead 20

Training Sample • Use the MCC11 DUNE “workspace” files • In future might be better to use isotropic particle guns with uniform energy distributions • Extract all track-like objects reconstructed from Pandora • Need to label my tracks with truth information. Currently matching by finding the true particle that contributes the most hits to the track (if there are more than one) • 0 = muon, 1 = charged pion, 2 = proton, (3 = kaon when I have more stats) • Total of 59128 tracks for training, and for 9522 testing • Trained for 5 epochs (only a few minutes on my Mac’s CPU) Leigh Whitehead 21

(Very) Preliminary Results 700 — All 
 — All 
 450 — Correct Prediction 600 — Correct Prediction 400 350 500 300 400 250 200 300 150 200 100 100 50 0 0 0 1 2 3 4 5 6 7 8 9 10 0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 Muon Energy (GeV) Pion Energy (GeV) Predicted Predicted Predicted 1 — All 
 Muon Pion Proton 1000 0.9 — Correct Prediction True 0.8 800 85.1% 14.0% 0.9% Muon 0.7 600 0.6 True Muon 
 18.3% 75.9% 5.8% 400 0.5 Pion Pion 
 200 0.4 Proton True 2.3% 10.6% 87.1% 0.3 0 Proton 0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 0 1 2 3 4 5 6 7 8 9 10 Proton Energy (GeV) Muon Energy (GeV) Leigh Whitehead 22