on behalf of the prospect collaboration andrea delgado, phd · open slide master to edit •siamese...
TRANSCRIPT
ORNL is managed by UT-Battelle, LLC for the US Department of Energy
Machine Learning Applications for Reactor Antineutrino Detection at PROSPECT
Andrea Delgado, PhD
On behalf of the PROSPECT collaboration
22 Open slide master to edit
Outline
● Overview of PROSPECT
● Antineutrino Event Reconstruction at PROSPECT
● Selected Machine Learning Applications at PROSPECT○ Positron ID through o-Ps tagging
○ Single PMT Event Reconstruction
○ Siamese LSTM for Pulse Pair Matching
○ MVA Analysis for Background Suppression
33 Open slide master to edit
Precision Reactor Oscillation and SPECTrum Experiment- Short-baseline reactor neutrino experiment.
Designed to address anomalies in flux and spectrum
- ~6% deficit in measurement of antineutrino.
- “Bump” around 5-7 MeV range
- Possible explanations:- Sterile neutrino hypothesis- Incorrect isotope predictions (235U)
- Challenge: Above-ground deployment - large backgrounds
7m
PROSPECT detector at High Flux Isotope Reactor (HFIR) at ORNL.
44cm
44 Open slide master to edit
Detector Design + Results- General description of experiment:
- The PROSPECT Reactor Antineutrino Experiment- Main published results:
- Measurement of the Antineutrino Spectrum from U 235 Fission at HFIR with PROSPECT
- First Search for short-baseline neutrino oscillations at HFIR with PROSPECT
- Most recent results just published.arxiv:2006.11210
Detector Performance- Detector response has been fully characterized.- PROSPECT-Geant4 (PG4): Geant4 - based simulation package for full
detector response simulation (by Michael Mendenhall)
Neutrino 2020 Talks & Posters - Talk: Recent Results from PROSPECT (B. Littlejohn) - June 25- Posters: 158, 516, 408, 527, 540, 556
55 Open slide master to edit
Antineutrino Event Reconstruction
Prompt signal: ~1-10 MeV positron energyDelayed signal: ~ 0.5 MeV neutron capture
Background suppression through set of cuts based on IBD topology, PSD, cosmic shower veto, etc...
Particle ID classification in 6Li-doped liquid scintillator
allows to differentiate ionization/nuclear
recoil/quenched n-Li
SIMULATION
Antineutrinos detected through Inverse Beta Decay (IBD) interaction
66 Open slide master to edit
Positron ID through o-Ps tagging- Can we ID a subset of positrons though
positronium formation?
- If the distortion in the timing distribution induced by o-Ps is not smeared by optical effects, we can use this feature as an extra handle for particle ID (P-ID).
- Preliminary results from optical simulation indicate that PROSPECT might be sensitive to these effects.
1ns Binning4ns Binning
3ns lifetime
20 ns lifetime
Para Positronium, (t=1, 2.5ns)
Ortho Positronium, (t=1,140ns)
PROSPECTSimulation
PROSPECTSimulation
PROSPECTSimulation
PROSPECTSimulation
77 Open slide master to edit
Positron ID through o-Ps tagging- Feed pulses from each detector into a Convolutional Neural Network.- For preliminary testing purposes, we are using 1ns sample widths, 150 samples taken from
simulated PMT pulses (PROSPECT uses 4ns digitizers)- Simulation details:
- 2 PMTs on each cell (11 x 14 cells) - 150 samples per pulse (1ns sample width)- Samples are 16-bit integers
- Map the 2 x 150 samples for each cell to the “channels” of a CNN - Output of CNN is a discriminant parameter (will be extended for generic particle ID soon)CNN Architecture
11x14x300 11x14x37 9x12x37 4x5x19 270 135 32 2
conv 1x1, 300
Stride 1Pad 0
conv 3x3, 32
Stride 1Pad 1
conv 3x3, 16
Stride 2Pad 0
flatten
Dense layers
88 Open slide master to edit
Positron ID through o-Ps tagging
CNN Architecture
11x14x300 11x14x37 9x12x37 4x5x19 270 135 32 2
conv 1x1, 300
Stride 1Pad 0
conv 3x3, 32
Stride 1Pad 1
conv 3x3, 16
Stride 2Pad 0
flatten
Dense layers
Model 3ns lifetime accuracy
20ns lifetime accuracy
Training time (CPU seconds)
XGBoost 0.5 0.55 860
CNN 0.5 0.64 9605
99 Open slide master to edit
Single PMT Event Reconstruction• Can we maintain PID performance while
using information from single PMT readout?1. Replicate the ML particle ID in double PMTs2. Train the model with single PMT info (from
double PMTs) ■ Compare the model prediction with
TRUE double PMT particle ID
Dense layers
Before bootstrap
After bootstrap
256 256 Output layer:
Label (4)
Input layer:E,dt,PE1,PE2,
PSD
Double-PMT
1010 Open slide master to edit
Single PMT Event Reconstruction• Can we maintain PID performance while
using information from single PMT readout?1. Replicate the ML particle ID in double PMTs2. Train the model with single PMT info (from
double PMTs) ■ Compare the model prediction with
TRUE double PMT particle ID
Dense layers
Single-PMT
256 256 Output layer:
Label (4)Input layer:
E,dt,PE1, PSD
1111 Open slide master to edit
Siamese LSTM for Pulse Matching• Cast pulse matching to a
“similarity” problem.• Keras-based implementation of
Siamese neural network.– Two identical subnetwork,
parameter updating is mirrored across both subnetworks
– Accuracy of ~73%.
Input 1:[E1, t1, x1]
Input 2 :[E2,t2,x2]
Prediction[0,1]
LSTM Encoder 1
Cosine distance
LSTM Encoder 2
Same weights, same model
How similar are the input vectors?
Predicted Label
True
Lab
el
1212 Open slide master to edit
MVA Analysis for Background SuppressionUsed in addition to set of rectangular selection cuts to optimize IBD selection efficiency.
• Keep signal rate, but reduce background contamination.• Use spatial and temporal correlations of prompt and
delayed signals.• Use ROOT’s TMVA package for simplicity + easy to add to
existing analysis framework.
MVA Classifiers
• Multi-Layer Perceptron (MLP)
• Boosted Decision Tree (BDT).
• Support Vector Machine (SVM).
Training/test datasets
• IBD simulation from PG4 for signal
• Reactor-off dataset for background
1313 Open slide master to edit
• Presented a few ideas on the directions our team is pursuing on ML applications to antineutrino data from PROSPECT experiment/simulation.
• Why Oak Ridge National Laboratory?
– Access to CPU/GPU resources through the Oak Ridge Leadership Computing Facility.
– Large community of machine learning experts.
– Synergy with existing groups working on machine learning + quantum computing (ask me about QC!)
• ML applications in PROSPECT can help enhance PSD and P-ID power and improve reactor antineutrino event reconstruction.
Summary
Questions?
Thank you!Andrea DelgadoAdriana GhiozziCorey GilbertBlaine HeffronXiaobin (Jeremy) LuDiego Venegas-VargasRosa Luz Zamora-PeinadoAlfredo Galindo-Uribarri
1515 Open slide master to edit
Quantum Computing Applications for Reactor Antineutrino Event Selection
- Earth mover's distance (EMD) is a measure of the distance between two probability distributions over a region D.
- Cast the minimization problem to an Ising Hamiltonian, which can then be solved in a quantum annealer, such as D-Wave.
The minimum cost of turning one pile of dirt into another