Development of an optimised event reconstruction for the Deep Underground Neutrino Experiment using machine learning and a multi-algorithm approach
Lead Research Organisation:
University of Warwick
Department Name: Physics
Abstract
Mousam will develop pattern-recognition algorithms in the Pandora framework[1], to reconstruct neutrino-induced events in the liquid-argon time-projection chambers (LArTPCs) to be deployed by the Deep Underground Neutrino Experiment (DUNE)[2]. DUNE is designed to help answer arguably the most important outstanding question of fundamental physics: "What is the origin of the matter-antimatter asymmetry in our Universe?". It will also provide precision measurements of the parameters governing neutrino oscillations and it has the potential to record a burst of neutrinos from a core-collapse supernova, providing a wealth of new information.
LArTPC pattern recognition is one of the most challenging problems in modern high energy physics and is a critical contribution to DUNE. LArTPCs provide "photograph quality" images of the charged particles produced in neutrino interactions. The images can be extremely complex, with a mix of overlapping track-like and shower-like topologies. Whilst the human brain can usually pick out the key features, it is a significant challenge to develop an automated, algorithmic solution. The pattern recognition is the single step in the DUNE workflow in which LArTPC images are examined in detail, so it is vital that information in the images is fully extracted.
The Pandora project champions a "multi-algorithm" approach to analysing LArTPC images, in which individual algorithms each look for specific features in event topologies. Many tens of algorithms carefully build up a picture of events and collectively provide a robust reconstruction. Pandora currently offers the most advanced and best documented LArTPC reconstruction and it is used extensively by the international neutrino physics community. Mousam will be working to develop novel pattern-recognition algorithms to identify features in events at DUNE.
In the first instance, Mousam will focus on understanding the current performance of the Pandora pattern recognition for DUNE events, identifying any weaknesses and designing algorithms to address any issues with specific topologies. Mousam will increasingly focus on the use of machine-learning approaches to drive the decisions made by pattern-recognition algorithms. He will develop algorithms that use machine-learning to classify individual hits in DUNE events as originating from track-like or shower-like particles. He will work to identify the positions of neutrino interaction vertices in DUNE events and perform the first studies to identify the vertices of secondary, downstream interactions. He will ensure the information extracted from machine-learning approaches is exploited effectively to drive a more performant pattern recognition.
Mousam will ultimately develop a physics analysis, using detailed knowledge of the pattern-recognition outputs to optimise selection of events and assess the sensitivity of DUNE to the parameters governing neutrino oscillations and/or CP violation in the neutrino sector.
[1] Eur. Phys. J. C (2018) 78: 82
[2] arXiv:1512.06148 [physics.ins-det]
LArTPC pattern recognition is one of the most challenging problems in modern high energy physics and is a critical contribution to DUNE. LArTPCs provide "photograph quality" images of the charged particles produced in neutrino interactions. The images can be extremely complex, with a mix of overlapping track-like and shower-like topologies. Whilst the human brain can usually pick out the key features, it is a significant challenge to develop an automated, algorithmic solution. The pattern recognition is the single step in the DUNE workflow in which LArTPC images are examined in detail, so it is vital that information in the images is fully extracted.
The Pandora project champions a "multi-algorithm" approach to analysing LArTPC images, in which individual algorithms each look for specific features in event topologies. Many tens of algorithms carefully build up a picture of events and collectively provide a robust reconstruction. Pandora currently offers the most advanced and best documented LArTPC reconstruction and it is used extensively by the international neutrino physics community. Mousam will be working to develop novel pattern-recognition algorithms to identify features in events at DUNE.
In the first instance, Mousam will focus on understanding the current performance of the Pandora pattern recognition for DUNE events, identifying any weaknesses and designing algorithms to address any issues with specific topologies. Mousam will increasingly focus on the use of machine-learning approaches to drive the decisions made by pattern-recognition algorithms. He will develop algorithms that use machine-learning to classify individual hits in DUNE events as originating from track-like or shower-like particles. He will work to identify the positions of neutrino interaction vertices in DUNE events and perform the first studies to identify the vertices of secondary, downstream interactions. He will ensure the information extracted from machine-learning approaches is exploited effectively to drive a more performant pattern recognition.
Mousam will ultimately develop a physics analysis, using detailed knowledge of the pattern-recognition outputs to optimise selection of events and assess the sensitivity of DUNE to the parameters governing neutrino oscillations and/or CP violation in the neutrino sector.
[1] Eur. Phys. J. C (2018) 78: 82
[2] arXiv:1512.06148 [physics.ins-det]
Studentship Projects
| Project Reference | Relationship | Related To | Start | End | Student Name |
|---|---|---|---|---|---|
| ST/S505857/1 | 01/10/2018 | 21/03/2023 | |||
| 2108560 | Studentship | ST/S505857/1 | 01/10/2018 | 21/03/2023 | Mousam Rai |