Search for low energy electron neutrinos in MicroBooNE using Deep Learning
This analysis is one of three searches for low energy electrons. For a summary of all three:
To read the results from the other two analysis:
An additional search looked for a photon-like excess:
What are we looking for?
Neutrinos are members of the Standard Model of particle physics. They come in three different types, or "flavors", which are distinguished by the particles they tend to produce when they interact. These flavors are known as the electron, muon, and tau neutrino. The 2015 Nobel Prize in physics was awarded for the observation of oscillations between different flavors, which implies the existence of tiny but nonzero neutrino masses.
One of the significant open puzzles for neutrino physicists concerns the Mini Booster Neutrino Experiment (MiniBooNE), a mineral oil Cherenkov detector that operated at Fermilab's Booster Neutrino Beam (BNB) from 2002 to 2017. MiniBooNE observes a significant excess of events in their detector that look like electron neutrino interactions, many more than one would expect from the Standard Model alone. However, the MiniBooNE detector has difficulty distinguishing between the electrons produced in electron neutrino interactions and photons produced by other processes.
Thus came MicroBooNE in 2007. The MicroBooNE detector is a liquid argon time projection chamber (LArTPC), which is particularly suited for disentangling the MiniBooNE puzzle. When charged particles from neutrino interactions pass through the argon in the detector, they leave behind a trail of ionized electrons. These electrons are drifted through an electric field to a series of three wire planes. Here the ionized electrons deposit charge on the wires, creating high-resolution images of the original charged particles. One can then distinguish between photons and electrons by leveraging small differences in the way they appear in these images.
Thanks to the picture-like quality, one can efficiently analyze LArTPC data through the use of image-based deep learning algorithms like convolutional neural networks. Such algorithms are used across a variety of fields, from self-driving cars to cat-species identification. Physicists from Columbia, IIT, MIT, SLAC and Tufts led the development of MicroBooNE's deep learning-based search for an excess of electron neutrinos in the BNB. This is the first use of such algorithms in a LArTPC to search for physics beyond the Standard Model. Presented here are the results of this search from the first half of MicroBooNE's full dataset.
What is a LArTPC?
The MicroBooNE detector is a 170-ton (85-active-ton) near-surface liquid argon time projection chamber (LArTPC) that collected data from 2015--2021 at Fermilab's Booster Neutrino Beamline. This makes it the first large-scale LArTPC to operate in the U.S.
When a neutrino interacts in the detector, many different particles can be produced. As charged particles from these interactions pass through the LAr medium, they leave behind a trail of prompt scintillation light and free ionization electrons. The scintillation light is collected by an array of photomultiplier tubes and can be used along with the beam trigger to fix the inititial time of the event. The ionization electrons are drifted through an electric field until they reach a series of three wire planes. The voltage bias on each plane is chosen such that the ionization electrons induce a voltage on wires in the first two planes before depositing their charge on wires in the final plane. The signals on each wire can then be used in conjucntion with the initial time to reconstruct the three-dimensional path of the original charged particle.
Animated illustration of a neutrino interaction in a LArTPC
Using Deep Learning
IMicroBooNE’s high-resolution data images provide the opportunity to use novel deep learning tools for neutrino event reconstruction. The results presented here make use of two deep convolutional neural networks as well as two boosted decision tree (BDT) ensembles used for 1mu1p and 1e1p event selection.
The first neural network, called SparseSSNet, identifies pixels in an image as five classes which are then condensed into two: track-like or shower-like -- for the LEE study.
This allows a series of traditional clustering algorithms to group pixels coming from the same particle together to reconstruct their kinematic properties.
SparseSSNet is a submanifold sparse convolutional neural network (CNN) and achieves an accuracy of >= 99% on the test sample. The sparse nature minimizes processing time (~0.5 seconds and image) and memory usage (~1GB).
Multi particle Identification
The second neural network currently used is MPID, a CNN used to identify the multiple particles in an event. MPID provides the probabilities that an interaction includes an e−, γ, µ−, π±, and protons. This represents the first use of CNNS to perform particle identification on real LArTPC data. The DL-LEE analysis uses this network to remove interactions with high γ scores from the background. MPID will continue to aid in future event selections.
"Traditional" Reconstruction Algorithms
A neutrino event vertex reconstruction is employed which finds "vee" shapes of either two track-like objects or one track-like and one shower-like object as labeled by SparseSSNet. These vertices form the basis of the event level reconstruction.
Reconstruction of track-like particles (protons and muons) begins at a vertex. 3D points are added to the track using an iterative stochastic search. The energy is calculated based on the length using the known stopping power of muons and protons
To reconstruct shower-like particles (photons and electrons), a triangle is placed and its parameters otimized within set limits so that the triangle encloses the maximum number of SparseSSNet labeled shower pixels. The energy of the particle is found by adding the charge of all of the shower pixels and converting to MeV. This process was verified using a sample of neutral pion decay photons and Michel electrons.
Boosted Decision Trees
A boosted decision tree (BDT) is a classification tool. Consisting on a sequence of binary decisions arranged in a tree structure, BDTs are a powerful pattern recognition tool that can be fit to datasets in order to discriminate sets within the data. We use BDTs in MicroBooNE to separate signal particle interactions from background using a variety of kinematic and topological variables describing the particle interaction in question. In this analysis specifically, we train an ensemble of BDTs using the XGBoost Python library.
Targeting a specific type of neutrino interaction
Identifying low-energy electron neutrino interactions among all background interactions is a formidable task. For every event with an electron neutrino interaction, there are hundreds of events with other types of neutrinos. Additionally, every event contains 20-30 cosmic rays that accumulate over the time it takes the ionization electrons to drift across the LArTPC. This analysis hones in on the signal by targeting a very specific type of electron neutrino interaction: events with one electron and one proton in the final state, where the particle kinematics are consistent with two-body scattering. These events typically come from charged-current quasi-elastic (CCQE) interactions, which dominate in the energy range relevant for this search.
The requirement that there is exactly one electron and one proton in the final state simplifies the final state topologies that we need to reconstruct. While some backgrounds can appear to have a similar topology due to misreconstruction, very few will also satisfy the kinematic criteria. We fully leverage the detailed information about both the electron and the proton from the MicroBooNE LArTPC to constrain the event. For example, this allows us to reconstruct the component of the momentum transverse to the neutrino beam, which is expected to be small in the case of true CCQE scattering but large in the case of most backgrounds. This and many other kinematic variables are used as inputs to the boosted decision trees.
This analysis does not observe an excess of electron neutrino charged-current quasielastic candidate events with one reconstructed proton and electron in the final state. This means that we can limit the fraction of these types of events contributing to the MiniBooNE anomaly to <38% with confidence of 95%. Analyses can see fluctuations, so it is good to have the result tested in several ways. Two other MicroBooNE analyses have also reported results. When all of the analyses are combined, we can say that the MiniBooNE result is not due to this type of interaction alone with a confidence of 99%. So what is MiniBooNE seeing? Stay tuned... MicroBooNE will next test if, along with a low level of these types of events, MiniBooNE may be seeing photons from some exotic source.
The Deep Learning Team
This work would not have been possible without the work of the entire MicroBooNE collaboration. The Deep Learning team comprises members from several institutions working as part the larger collaboration.
Gabriel Collin, Elizabeth Hall, Adrien Houlier, Jarrett Moon
Current Active Members
Professor of Physics, email@example.com
Graduate Student, firstname.lastname@example.org
Graduate Student, email@example.com
Current Active Members
Current Active Members
Current Active Members
Professor of Physics, firstname.lastname@example.org
Assistant Professor, email@example.com
Staff Scientist, firstname.lastname@example.org
Postdoctoral Researcher, email@example.com
Davio Cianci, Vic Genty
Rui An, Bryce Littlejohn
This work was sponsered by the Department of Energy and the National Science Foundation.
J. Conrad, N. Kamp, L. Yates. MIT Lunchtime Seminar. 2 November 2021