ACED: Physics-informed Geometric Deep Learning for Astrophysical Neutrino Reconstruction in IceCube DeepCore

About This Grant

Neutrinos are unique messengers, carrying information about the universe's most energetic astrophysical phenomena. Over the past decade, the IceCube Neutrino Observatory at the South Pole has made key discoveries by detecting high-energy neutrinos and identifying two active galaxies as neutrino sources. However, sub-TeV neutrinos (10–1000 GeV) remain a largely unexplored frontier with the potential to significantly expand our observation of the universe. This project leverages advanced artificial intelligence (AI) techniques to overcome computational challenges and improve the reconstruction of sub-TeV neutrinos using IceCube’s DeepCore subdetector. These advancements will enable detailed studies of astrophysical sources such as NGC 1068 at lower energy scales and pave the way for real-time public alerts of sub-TeV neutrinos, fostering coordinated follow-up observations across the electromagnetic spectrum. In addition to advancing neutrino astrophysics, the AI methodologies developed here will benefit a wide range of fields with similar challenges, including weather forecasting, neuroscience, smart cities, and precision farming, by enhancing the analysis of distributed sensor data. By integrating education initiatives, outreach programs, and undergraduate participation, this project promotes access to advanced research, supports STEM, and inspires the next generation of scientists and engineers. This project addresses the computational barriers to sub-TeV neutrino reconstruction through four interconnected research tasks. First, it designs a novel AI architecture capable of managing spatial and temporal irregularities in IceCube’s sensor data while embedding physics invariance for improved accuracy. Second, the project enhances computational efficiency by leveraging physics-informed inductive biases, enabling real-time processing of millions of neutrino events with low latency. Third, robust training methodologies will be implemented to address systematic uncertainties. Finally, the project investigates approximate symmetry modeling to allow AI models to adapt to practical deviations from the ideal physical model without compromising performance. These innovations will significantly improve the angular resolution and sensitivity of DeepCore to sub-TeV neutrinos, paving the way for transformative discoveries about astrophysical sources such as Seyfert galaxies as well as objects in the Milky Way. The tools and methodologies developed are designed to be broadly applicable, enabling breakthroughs in other scientific domains that rely on the analysis of complex spatiotemporal data. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.