Nue Energy Reconstruction with CNN Lars Hertel, Ilsoo Seong, - - PowerPoint PPT Presentation

Nue Energy Reconstruction with CNN

Lars Hertel, Ilsoo Seong, Jianming Bian 2018/08/20

Intro.

• Convolutional Neural Network (CNN) has been mostly used for classification
• This CNN has been implemented in the dunetpc software to classify neutrino

flavors and interactions (CVN)

• We use CNN to reconstruct Nue CC energy which is a regression problem : RegCVN
• We construct three pixel maps using Global Wire
• Use Nue sample of MCC10 for the training and evaluation:

prodgenie_nue_dune10kt_1x2x6_mcc10.0

• Implemented in the dunetpc using Tensorflow and tested with dunetpc v07_00_01

(not committed yet)

2

Event Selection and Global Wire

• Select Nue CC MC events: No hits near the edge of the FD
• Global Wire

○ modified the GlobalWire algorithm in the “BlurredClusteringAlg” module

○ It shows a clean and continuous event from one image.

Global Wire Local Wire using reco Hit Hits on two TPCs Pass through APA

3

Pixel Map

• Three input pixel maps: U, V, and Z planes
• Pixel map size: 280x400 (actual covered space: 1680 ticks x 400 wires) → 6 ticks are merged
• Use ADC counts and TDC units from Wire instead of using reconstructed Hit

4

Training Details

Stochastic gradient descent (SGD): Use SGD variant ADAM [Kingma et al. 2014] with 1e-3 learning rate. Batch size n: 16 Loss: Training over 460,000 examples for approximately 20 epochs.

Inception Max Pool 3xConv2d Max Pool

Concat

5 3xConv2d Max Pool 3xConv2d Max Pool Inception Max Pool Inception Max Pool Inception Max Pool Inception Max Pool Inception Max Pool Inception Average Pool

Energy

Plane U Plane V Plane Z

NueCC Energy Resolution

• Black line: energy resolution from RegCVN
• Blue line: energy resolution from dune-reco
• RegCVN has less bias and better resolution
• Fit with Gaussian near the peak region
• Sigma of RegCVN: 5.2% , Std. : 8.0%

6

Energy Resolution vs True Energy

• Mean and RMS of energy resolution
• RegCVN shows less bias than and smaller RMS
• Need improvement in the low energy region

7

Summary and Plan

• RegCVN promises better Nue energy resolution
• The RegCVN module is implemented in the dune art framework and ready to commit
• To improve the resolution in the low energy region, we will up-weight low energy

examples during training

• RegCVN will be also used to reconstruct electron shower energy in Nue events

8