Session T15: Computational Physics and Algorithms II

Single Particle Identification with a Context-Enriched Convolutional Neural Network in the NOνA Experiment

In 2016, NOνA was the first HEP experiment to employ a convolutional neural network (CNN) in a physics result, using the CNN to classify neutrino events. However, the physics analyses performed by NOνA require further identification and reconstruction of particles in the interaction final states. We have developed the first implementation of a CNN for single particle classification which employs context-enhanced inputs. Using contextual information from the neutrino interaction that produces the particles provides additional information to the training, extending the capabilities of our original classifier. This implementation uses a four-tower siamese architecture for separation of independent inputs and inclusion of contextual information. This classifier distinguishes between electrons, muons, photons, pions, and protons with a global efficiency and purity of 83.7% and 83.5%, respectively. In this talk I will describe our implementation of NOνA's single particle CNN, discuss the advantages of adding context information, and provide case-studies of the applications of the classifier.   

Neutrino Energy Reconstruction with Regression Convolutional Neural Network at DUNE

The study of neutrino oscillations is one of the primary physics goals of the DUNE experiment. The neutrino oscillations are functions of neutrino energies, hence the reconstruction of neutrino energies with high resolution is important to accomplish the successful measurements. We had developed a method to reconstruct energy taking the sum of the reconstructed track or shower and hadronic energies. This method can be limited because of the complicated event topology, low hadronic energy resolution, and invisible energy. Thus, we developed regression Convolutional Neural Networks (CNNs) to estimate electron neutrino energy with deconvoluted waveform inputs. Compared with the kinematics-based reconstruction, this method shows a significantly better energy resolution. In this talk, I will describe the methods to reconstruct neutrino energies with kinematic-based and regression CNN-based reconstructions at DUNE.   

Imaging Electrons in ATLAS

Efficient and accurate electron reconstruction, identification, and calibration are critical for signal selection and uncertainty reduction in a broad range of ATLAS analyses. Traditionally, electron algorithms are built using physics-motivated, derived variables. This talk explores an alternate method for representing electrons by building images using calorimeter cells. These images can then be processed using machine learning algorithms like convolutional neural networks, yielding improved identification and calibration.   

GEARS - a Geant4 Example Application with Rich features and Small footprints

We'd like to introduce GEARS, a Geant4 Example Application with Rich features yet Small footprint. The entire C++ coding is minimized down to a single file with about 600 SLOC. This is achieved mainly by utilizing Geant4 plain text geometry description, build-in UI commands (macros), C++ inheritance, and unconventional uses of some standard Geant4 functionalities. It is ideal for student training and fast implementation of small to medium-sized experiments.   

Neural Network Algorithms for the CMS Level-1 Muon Trigger

The CMS trigger system selects interesting events from collision events in the LHC to keep the stored data to manageable size. The level-1 trigger is custom electronics designed to trigger events online very quickly and efficiently to meet tight timing requirements of microseconds to reach a decision. The current endcap muon trigger uses an external memory look-up table for the momentum assignment, and a trained Boosted Decision Tree for the algorithm. In preparation for the High Luminosity LHC, further reduction in the trigger rate is necessary in order to maintain similar thresholds in harsher conditions (higher pile-up). A more accurate assignment of the transverse momentum of muons reduces the rate coming from mismeasured lower momentum muons. Neural Networks can use a larger number of features to improve the momentum assignment and implementing neural nets directly in the FPGA logic without external memory frees us from the limitation on the number of address bits for a look-up table. I have investigated optimizing existing Neural Networks architecture. I’ll show performances for the pT resolution and the trigger rate and the optimization result for the architecture model including a robustness study of the algorithm.