Session Y13: Computational Techniques For Event Reconstruction - II

Live

The Deep Underground Neutrino Experiment (DUNE) will use the liquid argon time projection chamber (LArTPC) technology. LArTPC detectors can collect high-resolution data of charged particles' trajectories. An example of this type of detector is ProtoDUNE-SP that is the prototype of the single-phase DUNE far detector using full-scale components and a charged-particle beam that allows measuring the detector's calorimetric response to hadronic particles and electromagnetic showers. The Convolutional Neural Networks has been developed and employed in the analysis of scientific data from the protoDUNE detector, which exploit the advantages of a liquid argon time projection chamber (LArTPC). Despite the high-resolution images and the fine details that the detector can capture, the classification of the different types of particles and interactions is still a challenge. With this motivation, we present the details of an algorithm capable of generating a pixel-level label for supervised training using Sparse Convolutional Neural Networks.   

Live

The IceCube Neutrino Observatory is the world's largest neutrino telescope, located under the ice in the South Pole. It aims to detect astrophysical and atmospheric neutrinos to discover cosmic sources of neutrinos and to better constrain fundamental neutrino parameters, such as the mixing parameters controlling neutrino flavor oscillations. IceCube's 3D hexagonal array of 5160 digital optical modules (DOMs) detects light from neutrino interactions in the ice and can be used to reconstruct the incident neutrino's energy, direction, etc. A convolutional neural network, typically used for image identification, can be adapted to IceCube's DOM array to reconstruct properties of the incident neutrino. Neural networks have been successfully used in IceCube for reconstruction at higher energies (100 GeV - 10 PeV), but understanding fundamental neutrino parameters requires exploring neutrino events at lower energies (less than 100 GeV). These events leave sparse signals in IceCube, so the network requires reoptimization. In this talk, I will present my work applying a convolutional neural network to reconstruct the energy and direction of low energy neutrino events in the IceCube detector.   

Live

NOvA is a long-baseline neutrino oscillation experiment that is designed to probe the neutrino mass hierarchy and mixing structure by looking for a $\nu_{e}$ ($\bar{\nu}_{e}$) appearance signal. It uses two functionally identical liquid scintillator detectors $14.6$ mrad off-axis from the NuMI beamline at Fermilab, allowing for a tightly focused $\nu_{\mu}$ flux peaked at around 2 GeV. In order to make oscillation parameter measurements with high precision, it is important to reconstruct neutrino energies with good resolution as the oscillation probability is a function of neutrino energy. This is not straightforward due to complicated event topologies and large uncertainties on the underlying interaction models. To address this, NOvA has developed a deep learning based CNN that is able to estimate $\nu_{e}$ energies non-parametrically. This approach not only gives superior energy resolutions to traditional kinematic-based estimations, but also shows better behavior under changes to the interaction model; thus enabling us to reduce systematic uncertainties on the final measurement. In this talk, I shall present a summary of the CNN approach and highlight its response to the underlying physical model.   

Live

NOvA is a long-baseline neutrino oscillation experiment designed to measure the muon (anti)neutrino disappearance and electron (anti)neutrino appearance in the off-axis Fermilab NuMI beam. It uses two functionally identical detectors separated by 810km and a narrow band beam centered around 2GeV. Neutral pions are a significant background to the electron (anti)neutrino appearance measurement as the photons coming from neutral pion decay may be misidentified as an (anti)neutrino appearance signal. The high statistics antineutrino mode data in the Near Detector (ND) can be used to perform a measurement of the cross-section of the muon antineutrino charged-current (CC) neutral-pion production in the resonance regime. The analysis will use a convolutional visual network (CVN) trained on simulated particles to identify neutral pions in the final state. The status of the analysis and preliminary sensitivities will be presented.   

Live

Several interesting BSM models predict the possibility of low mass resonances. But the kinematic thresholds used in the current set of triggers make CMS blind to these resonances. To overcome this problem CMS has implemented Data Scouting techniques that allow trigger thresholds to be lowered by saving a very limited amount of trigger-level event information offline. Here we present the searches that used this data scouting technique to set some of the strongest limits to date for low mass resonances in multi-jet and di-muon channels. I will talk about the various new techniques we developed to use the scouting dataset to search for low mass boosted hadronic resonances and ultra-low mass long-lived di-muon resonances. I will also talk about the new fitting techniques that are being used for background prediction at these low masses regimes.   

Live

One of the most common applications of machine learning in high energy physics is in event selection (and categorization). The physics goals of event selection and categorization are to improve the significance of a potential excess (for signal discovery/upper limit setting analyses), and to reduce the uncertainty of a parameter measurement (parameter measurement analyses). Event selection using machine learning is based on the "signal is better than background" heuristic. While it is clear how the heuristic would help with the physics goals, it turns out that they are not completely aligned. In fact, certain signal events could be worse for the sensitivity of an analyses than certain background events. In this talk we will provide optimal event selector and categorizer training prescriptions designed to maximize the expected statistical significance of an excess (by changing how ML outputs are used), and minimize the statistical uncertainty of a measurement (by changing the supervisory signal used in training the ML algorithms). Along the way, we will point out exactly how our methods realign the goals of event selection and categorization with the physics goals.   

Live

The relevant information from collision events from the Large Hadron Collider (LHC) and other colliders can be represented as spatial point data in a suitable phase space. The observation of sharp discontinuities in the observed event number density would hint at the presence of new physics beyond the Standard Model. We apply and further improve upon some known wombling techniques from other fields. We illustrate our method with simulated high energy data.   

On Demand

Use of machine learning (ML) models to classify signal/background is a critical component of many analyses in particle physics and astrophysics. In these applications, maximizing the statistical significance ($\sigma$) of the resulting signal sample is often paramount to maximizing classification accuracy, the typical figure of merit for model optimization. However, accuracy and statistical significance can in fact be in tension in applications where signal and background are inseparable in some regions of the input feature space. We present a novel approach to multi-class problems that optimizes a neural network to predict into an expanded category space using a loss function that combines accuracy and statistical significance. This approach allows the model to ignore regions of the input space in which events from multiple classes are impossible to separate without overfitting. We demonstrate the application of the relegation approach to hadronic-physics datasets and show that it produces analysis significance comparable to logistic regression.   

Not Participating

In this talk we discuss application of the recurrent neural networks to the task of energy reconstruction at the NOvA experiment. NOvA is a long-baseline accelerator based neutrino oscillation experiment that holds a leading measurement of the $\Delta m_{32}^2$ oscillation parameter. In order to achieve good estimation of the oscillation parameters it is imperative to have a good neutrino energy estimation algorithm. We have developed a new energy estimation algorithm that is based on a recurrent neural network. The new energy estimator has better performance than the previous NOvA energy estimation algorithm, and it is less affected by some of the major NOvA systematics. Using this new energy estimator has potential to significantly improve NOvA sensitivity to the oscillation parameters.