Speaker
Description
The IceCube Neutrino Observatory’s DeepCore sub-array lowers the energy
detection threshold to the GeV scale, opening the sub-TeV window for studying
potential astrophysical neutrino sources, such as NGC 1068, a source previously
associated with high energy neutrinos by IceCube and expected to emit neutrinos
down to sub-TeV energy. However, reconstructing events in the sub-TeV range is
challenging due to sparse photon statistics and background noise. Likelihoodbased algorithms are computationally expensive, requiring approximately 15
minutes per event, and are often limited in resolution.
To address these challenges, we present TANGO (Transformer ANgles for GrecO), a
deep learning model optimized for the GRECO (GeV Reconstructed Events with
Containment for Oscillations) dataset, which selects DeepCore events for
astrophysical neutrino studies covering the transition from GeV to TeV energies.
TANGO innovatively combines Graph Neural Networks (EdgeConv) with Transformer
layers to capture both local feature extraction and global long-range dependencies
among detector pulses in its architecture. A key feature of the model is the joint
reconstruction of the neutrino direction and interaction vertex using a weighted
multi-task loss function, which leverages the correlation between spatial and
directional features.
TANGO is trained and evaluated on the simulated GRECO dataset. Performance
evaluations demonstrate that TANGO significantly outperforms likelihood methods
in both accuracy and efficiency. The model reduces inference time to approximately
10 milliseconds per event. In terms of reconstruction quality, TANGO improves
angular resolution by a factor of 1.2 to 2.5. Vertex position reconstruction shows
even more dramatic gains: the vertical depth resolution improves by a factor of 1.7
to 2.5, while the radial distance resolution improves by a factor of 5 to 10.
Furthermore, TANGO provides robust angular uncertainty estimations, correcting
systematic biases often found in traditional reconstruction. These advancements
enhance IceCube’s sensitivity to sub-TeV sources and provide a validated
reconstruction framework for the upcoming IceCube Upgrade.