AI in the
Dark Matter Experiment DEAP-3600
Antonio Giménez Alcázar on behalf
CIEMAT DM Group & DEAP collaboration
XIV CPAN Days
23-25 November 2022 Bilbao
1. DEAP-3600 Experiment
— Experiment— Detector
— Backgrounds
• DEAP-3600
is a single-phase liquid argon detector designed for the WIMP search. It is placed in SNOLAB, a deep underground laboratory which uses the 2 km of rock overburden to filter out cosmic-ray muons.Dark matter Experiment using Argon Pulse-shape Discrimination
DEAP-3600 is the largest runningLAr detector designed for a dark matter search
Phys. Rev. D 100, 022004 (2019)
• WIMP dark matter search
• Profile likelihood ratio analysis
• Blind analysis on full dataset
• Other searches
• 5.5 MeV solar axions
• 8B neutrino absorption signal (inverse beta decay)
• 39Ar Measurements
• 39Ar specific activity
• 39Ar half-life
• 39Ar decay spectrum and nuclear parameters
• Muon veto instrumentation and muon flux measurement
• Muon flux annual modulation (or absence thereof)
DEAP-3600 prospect analysis
Dark matter Experiment using Argon Pulse-shape discrimination
Current Picture
• WIMP dark matter search
• Profile likelihood ratio analysis
• Blind analysis on full dataset
• Other searches
• 5.5 MeV solar axions
• 8B neutrino absorption signal (inverse beta decay)
• 39Ar Measurements
• 39Ar specific activity
• 39Ar half-life
• 39Ar decay spectrum and nuclear parameters
• Muon veto instrumentation and muon flux measurement
• Muon flux annual modulation (or absence thereof)
DEAP-3600 prospect analysis
• DEAP-3600
is a single-phase liquid argon detector designed for the WIMP search. It is placed in SNOLAB, a deep underground laboratory which uses the 2 km of rock overburden to filter out cosmic-ray muons.The Detector
• Ultraclean acrylic vessel filled with 3.3 tonnes of liquid argon. The inner surface is coated with TPB to shift the UV light generated by argon scintillation to the visible region, which can then be
transmitted to the 255 photomultiplier tubes (PMTs).
• PMTs cover 75% of the sphere. Grouped in 35 rings of 5 or 10 PMTs.
• 48 veto PMTs looking outwards.
Astroparticle Physics 108 (2019) 1-23
What happens inside?
LAr as target Gas
1) Dark matter elastically scatters off argon nuclei
Ar
χ
What happens inside?
LAr as target Gas
1) Dark matter elastically scatters off argon nuclei
Ar
χ 2) Excited/ionized atoms form excited dimer states
and relax to ground state via scintillation of 128 nm photons (VUV).
Denver Whittington, Syracuse University
What happens inside?
LAr as target Gas
1) Dark matter elastically scatters off argon nuclei
Ar
χ
3) Scintillation photons pass through TPB wavelength shifter, become in the range of420 nm photons (visible).
2) Excited/ionized atoms form excited dimer states and relax to ground state via scintillation of 128 nm photons.
LAr Spectra
TPB Spectra
V.V. Golovko, et al.
What happens inside?
LAr as target Gas
1) Dark matter elastically scatters off argon nuclei
Ar
χ
3) Scintillation photons pass through TPB wavelength shifter, become 420 nm photons.
4) Photons collected by light guides, detected by the 255 PMTs.
Inspired on Joseph McLaughlin slides for ICHEP 2022
2) Excited/ionized atoms form excited dimer states and relax to ground state via scintillation of 128 nm photons.
Pulse shape discrimination (PSD)
Ar39 & γ
α
decays ROIPulse shape allows us to discriminate against γ and β backgrounds.
Singlet state-fast decay
Triplet state-slow decay
Eur. Phys. J. C 81,823 (2021)
Backgrounds
Simulated relationship in reconstructed z vs. PE for α-decays
• High-energy alpha decays observed from the liquid argon volume are well-described by our background model.
• Dominant background in WIMP region due to neck alphas.
Phys. Rev. D 100, 072009
Dominant Background: neck alphas
Phys. Rev. D 100, 022004
• The largest contribution to background are events generated by α particles.
• The origin is on the flow guides (FG). These surfaces
contain radioactive isotopes from the Radon chain which emit α-particles that, in contact with the LAr condensed on the guide surfaces, emit ultraviolet radiation.
• α particles generate up to 5000 photoelectrons (PE) in the PMTs. FG are not coted in TPB, and most of the UV photons are absorbed in the acrylic. This results in
shadowed event topologies in which only a small fraction of the emitted photons reach the PMT.
Dominant Background: dust alphas
• Alpha decays from trace amounts of dust particulates in LAr create low-PE events originating in the centre of the detector.
2. Deep Learning
— Recovering signal sensitivity— Explainable AI
Recovering Signal Sensitivity with NN
• Ongoing MVA/machine learning analysis, with improved signal
acceptance and lower backgrounds. Three different algorithms are used: RF, XGB and NN (CIEMAT)
• Convolutional neural networks (CNNs) are a specialized kind of neural network for processing data that have a known pattern.
• In our case, we are interested in 2D CNNs that can identify a topological image of a WIMP event.
2D CNN takes advantage of spatial correlations of PMTs
Light Pattern = # Photoelectrons in each PMT
Time Pattern = Time of first photon to arrive to each PMT
Neck alphas
• We develop Machine Learning models able to classify neck events (background) vs nuclear recoils (signal).
• DEAP collaboration requests a neck-event rejection of 99.9%.
• We define acceptance as: fraction of signal events surviving the neck-event rejection criteria.
Explainable AI: GRAD-CAM
• GRAD-CAM produces a coarse localization map highlighting the important regionsin the image for predicting the concept.
• Ramprasaath R., et, al (2019), Grad-CAM: Visual Explanations from Deep Networks via Gradient-based Localization.
Why does it predict what it predicts?
• SHAP calculates individual explanations for each event. It explains the difference between the output value and the expected value.
S. M. Lundberg & Su-In Lee NIPS 2017
Explainable AI: SHAP
Top of detector Bottom of detector
4. Conclusions
• Sensitivity currently limited by neck and dust alpha backgrounds.
• Deep learning techniques help to recover design sensitivity.
• XAI methods are helpful to
understanding how predictions are made.