Neutrino interaction classification with the DUNE Convolutional - - PowerPoint PPT Presentation

Neutrino interaction classification with the DUNE Convolutional Visual Network Leigh Whitehead, Saúl Alonso- Monsalve

DUNE Collaboration Call 3 April 2020

Leigh Whitehead, Saúl Alonso-Monsalve

Purpose of the paper

2

• Paper of the CVN Particle ID used in the TDR sensitivities.
• Title: “Neutrino interaction classification with the DUNE Convolutional

Visual Network.”

• Primary authors: Leigh Whitehead, Saúl Alonso Monsalve.
• ARC: Alex Himmel, Andy Blake, Dan Cherdack, Andrea Scarpelli, Taritree

Wongjirad.

• DUNE-doc-14125.
• Target journal: Physical Review D (PRD).
• Deadline of the review: Monday, April 13th.

Leigh Whitehead, Saúl Alonso-Monsalve

Overview

3

• 1. Introduction to DUNE.
• CP-violation measurement.
• DUNE FD simulation and reconstruction.
• 2. CVN neutrino interaction classifier.
• CVN inputs, outputs, and network architecture.
• Training details.
• 3. Neutrino flavor identification performance.
• Focus on CC νe and CC νμ selections.
• Efficiencies of 90% for CC νe and 95% for CC νμ.
• 4. Exclusive final state results.
• Results using the CVN outputs that count the number of final-state particles for:

protons, charged pions, and neutral pions.

• 5. Robustness.
• Evaluate the CVN performance as a function of different observable physics

parameters.

• 6. Conclusion.

Leigh Whitehead, Saúl Alonso-Monsalve

Introduction to DUNE

4

• Gives an introduction to neutrino physics, the experiment and the TDR

CP-violation analysis.

• We include two plots of the event selection from the TDR.
• These are the CC ve

and CC ve events for a range of dCP values.

Leigh Whitehead, Saúl Alonso-Monsalve

CVN neutrino interaction classifier

5

• This section describes the details of the CVN.
• The architecture including all of the inputs and outputs.
• Details of the training sample and methods.
• Example input images of signal and background events.

Leigh Whitehead, Saúl Alonso-Monsalve

CVN neutrino interaction classifier

6

• We include a schematic diagram of the

network architecture.

• Provide full details of the training

procedure and the samples used.

• Plots of the loss and accuracy

during the training process.

Leigh Whitehead, Saúl Alonso-Monsalve

Neutrino flavor identification performance

7

• This section describes the main result of the paper.
• Corresponds directly to the results shown in the TDR.
• We show the distribution of the CVN classifier score for the ve and vu

hypotheses for FHC and RHC beams.

Leigh Whitehead, Saúl Alonso-Monsalve

Neutrino flavor identification performance

8

• This section describes the main result of the paper.
• Corresponds directly to the results shown in the TDR.
• We show the distribution of the CVN classifier score for the ve and vu

hypotheses for FHC and RHC beams.

Leigh Whitehead, Saúl Alonso-Monsalve

Neutrino flavor identification performance

9

• The “final result” shown in this paper is the selection efficiency using

the CVN compared to the CDR analysis.

Leigh Whitehead, Saúl Alonso-Monsalve

Neutrino flavor identification performance

10

• The “final result” shown in this paper is the selection efficiency using

the CVN compared to the CDR analysis.

Leigh Whitehead, Saúl Alonso-Monsalve

Exclusive final state results

11

• This section describes the potential of the other CVN outputs that

count the number of final state particles.

• This goes beyond what was used in the TDR analysis.
• Provides a clear proof-of-principle for sub-dividing the FD event

selection in the future.

Leigh Whitehead, Saúl Alonso-Monsalve

Robustness

12

• This section is aimed at convincing the reader that we understand what

the CVN is doing and that it behaves in an expected way as a function

• f different physics parameters.
• For example, we see that the selection efficiency drops as a we

increase the hadronic energy in the system.

Leigh Whitehead, Saúl Alonso-Monsalve

Robustness

13

• This section is aimed at convincing the reader that we understand what

the CVN is doing and that it behaves in an expected way as a function

• f different physics parameters.
• Similarly, we see NC background events with higher acceptance as the

pion energy increases.

Leigh Whitehead, Saúl Alonso-Monsalve

Conclusion

14

• Reiterate the impressive performance of the CVN and that it is a robust

classification algorithm.

• Suggest, dependent on further studies, that the particle counting
• utputs have the potential to increase the sensitivity further.

Leigh Whitehead, Saúl Alonso-Monsalve

Public data release

15

• Gitlab project:
• https://gitlab.cern.ch/salonsom/cvn-paper.
• The code runs the CVN over a small sample.
• The sample consists of 20 random MC events.
• There is a README file available with a detailed description of the code.
• The testing script produces a file called results.txt. This can be compared

to ./output/expected_results.txt to ensure the code has executed correctly.

Mechanics of the review

Leigh Whitehead, Saúl Alonso-Monsalve 16

• We’re currently about half-way through the review period.
• Comments are due back on Monday, April 13th.
• Send to ahimmel@fnal.gov, leigh.howard.whitehead@cern.ch,

saul.alonso.monsalve@cern.ch.

• In doc-14125 you can find a draft of the paper as well as a spreadsheet

template for comments.

• Please fill in comments in the appropriate tab depending on what

the comments is on along with the column marking line, fig #, etc.

• Put an X in the “Minor?” column if you don’t feel you need a

response to your comment from the authors.

Neutrino interaction classification with the DUNE Convolutional Visual Network Leigh Whitehead, Saúl Alonso- Monsalve

DUNE Collaboration Call 3 April 2020