Event reconstruction using Graph Neural Networks in Project 8
ORAL
Abstract
Project 8 seeks to determine the absolute neutrino mass by measuring the endpoint of the Tritium β-decay spectrum using Cyclotron Radiation Emission Spectroscopy (CRES).
A CRES event manifests in the time-frequency representation, such as short term Fourier transform, as a series of tracks, each corresponding to electron’s motion between scatters. To accurately reconstruct the kinematic properties of the β-decay electrons, it’s important to identify the individual track properties in noisy time-frequency CRES data. We present a machine-learning approach using Graph Neural Networks (GNNs) for event reconstruction in Project 8. GNNs are particularly advantageous because they handle variable-sized inputs, naturally model sparse detector data, and maintain permutation invariance. In this framework, time–frequency bins are represented as graph nodes that are connected based on their temporal and spectral properties. This graph representation allows the network to exploit the intrinsic topology and the underlying physics of the data to perform track identification without requiring explicit feature engineering. We will present the performance of GNN models trained on simulated CRES data, and discuss a roadmap to perform event reconstruction on upcoming experimental datasets using these models.
A CRES event manifests in the time-frequency representation, such as short term Fourier transform, as a series of tracks, each corresponding to electron’s motion between scatters. To accurately reconstruct the kinematic properties of the β-decay electrons, it’s important to identify the individual track properties in noisy time-frequency CRES data. We present a machine-learning approach using Graph Neural Networks (GNNs) for event reconstruction in Project 8. GNNs are particularly advantageous because they handle variable-sized inputs, naturally model sparse detector data, and maintain permutation invariance. In this framework, time–frequency bins are represented as graph nodes that are connected based on their temporal and spectral properties. This graph representation allows the network to exploit the intrinsic topology and the underlying physics of the data to perform track identification without requiring explicit feature engineering. We will present the performance of GNN models trained on simulated CRES data, and discuss a roadmap to perform event reconstruction on upcoming experimental datasets using these models.
*This work is supported by the US DOE Office of Nuclear Physics, the US NSF, the PRISMA+ Cluster of Excellence at the University of Mainz, and internal investments at all institutions.
–
Presenters
-
Vivek Sharma
- University of Pittsburgh