Periodic Reporting for period 1 - GraphNeT (Graph convolutional neural networks for neutrino telescopes)
Reporting period: 2021-09-01 to 2023-08-31
Neutrinos, an elusive particles that comes in three flavours, can change their flavour as they travel through space, a phenomenon called neutrino oscillation. This unique characteristic sets them apart from other particles, and additionally implies that neutrinos are massive. Despite their significance, the neutrino sector of the Standard Model of particle physics remains poorly understood. The fact that neutrinos have mass, for instance, unlike what the Standard Model predicts, hints at physics beyond our current understanding.
Neutrino oscillation experiments at the IceCube Neutrino Observatory may help answer this and other big scientific questions like it. However, a major challenge lies in the technical task of efficiently identifying and reconstructing neutrino interaction events in the vast stream of data recorded by IceCube. Existing algorithms for doing so are prohibitively slow and are not compatible with planned detector upgrades.
To tackle this challenge, this project aimed to develop novel reconstruction algorithms using machine learning (ML), particularly graph neural networks (GNNs). These algorithms could significantly improve event reconstruction in IceCube and similar experiments and thereby provide better and more accurate data for physics analysis.
Despite the project ending early, it successfully showcased the potential of ML and GNNs to enhance event reconstruction in IceCube, surpassing existing algorithms in both accuracy and speed. This breakthrough paves the way for addressing pressing questions in the neutrino sector and beyond.
Neutrino oscillation experiments at the IceCube Neutrino Observatory may help answer this and other big scientific questions like it. However, a major challenge lies in the technical task of efficiently identifying and reconstructing neutrino interaction events in the vast stream of data recorded by IceCube. Existing algorithms for doing so are prohibitively slow and are not compatible with planned detector upgrades.
To tackle this challenge, this project aimed to develop novel reconstruction algorithms using machine learning (ML), particularly graph neural networks (GNNs). These algorithms could significantly improve event reconstruction in IceCube and similar experiments and thereby provide better and more accurate data for physics analysis.
Despite the project ending early, it successfully showcased the potential of ML and GNNs to enhance event reconstruction in IceCube, surpassing existing algorithms in both accuracy and speed. This breakthrough paves the way for addressing pressing questions in the neutrino sector and beyond.
During the project, the primary focus was on creating GraphNeT, an open-source software framework for building, training, and using GNNs in neutrino telescope event reconstruction. This framework serves as a valuable resource for both ML developers and physicists, offering state-of-the-art GNN algorithms for neutrino research. GraphNeT's goal is to provide a user-friendly, versatile, reusable package following best software engineering practices, enabling physicists to access powerful tools for their scientific endeavours.
The project involved leading a group of developers from across IceCube and overseeing the design, development, and maintenance of the framework. Regular presentations at IceCube working group and collaboration meetings were made to share progress with the collaboration. As a result, GraphNeT has already been utilised in IceCube publications and has been proven to enhance performance on various physics reconstruction tasks by over 20%. Furthermore, GNNs built with GraphNeT are fast enough to analyse all data from IceCube in real-time at the South Pole.
Beyond software development, considerable efforts were dedicated to personal development, meticulous project management, and coordination. Weekly GraphNeT developer meetings were organised to facilitate collaboration, and several students were co-supervised in GNN-related projects for IceCube. Seminars on ML applications in neutrino telescope experiments were hosted, and a well-attended in-person workshop gathered participants from various physics experiments to expand the project's impact and ensure its continued exploitation beyond the project's conclusion.
The project involved leading a group of developers from across IceCube and overseeing the design, development, and maintenance of the framework. Regular presentations at IceCube working group and collaboration meetings were made to share progress with the collaboration. As a result, GraphNeT has already been utilised in IceCube publications and has been proven to enhance performance on various physics reconstruction tasks by over 20%. Furthermore, GNNs built with GraphNeT are fast enough to analyse all data from IceCube in real-time at the South Pole.
Beyond software development, considerable efforts were dedicated to personal development, meticulous project management, and coordination. Weekly GraphNeT developer meetings were organised to facilitate collaboration, and several students were co-supervised in GNN-related projects for IceCube. Seminars on ML applications in neutrino telescope experiments were hosted, and a well-attended in-person workshop gathered participants from various physics experiments to expand the project's impact and ensure its continued exploitation beyond the project's conclusion.
This project has made significant strides in exploring the potential of GNNs for event reconstruction at neutrino telescope experiments, surpassing the previous state-of-the-art techniques. ML, particularly GNNs, has proven to be highly suitable for enhancing reconstruction fidelity across various physics tasks, while drastically reducing inference time. Additionally, the collaborative approach to developing open-source ML software packages has shown potential for broader impact compared to the fragmented development efforts that are more common at physics experiments.
This progress is just the beginning. The initial performance studies focused on isolated physics tasks based on datasets that were selected based on simpler, existing algorithms. However, GNNs have the potential to deliver improvements throughout the entire physics analysis chain. From triggering and real-time filtering at the South Pole to event selection and final, high-precision reconstruction of physics quantities, GNNs could have a substantial cumulative effect, far exceeding the promising initial results.
Looking ahead, GNNs like those developed through GraphNeT, are well-positioned to have a profound and far-reaching impact at neutrino telescope experiments and beyond. Similarly, the collaborative and open-source approach employed in this project can help foster knowledge sharing and accelerate progress across multiple disciplines. Combined, these results of the project can help accelerate scientific discoveries and our understanding of some of the most fundamental properties of the Universe.
This progress is just the beginning. The initial performance studies focused on isolated physics tasks based on datasets that were selected based on simpler, existing algorithms. However, GNNs have the potential to deliver improvements throughout the entire physics analysis chain. From triggering and real-time filtering at the South Pole to event selection and final, high-precision reconstruction of physics quantities, GNNs could have a substantial cumulative effect, far exceeding the promising initial results.
Looking ahead, GNNs like those developed through GraphNeT, are well-positioned to have a profound and far-reaching impact at neutrino telescope experiments and beyond. Similarly, the collaborative and open-source approach employed in this project can help foster knowledge sharing and accelerate progress across multiple disciplines. Combined, these results of the project can help accelerate scientific discoveries and our understanding of some of the most fundamental properties of the Universe.