Machine Learning and Reconstruction

Our group develops machine-learning methods for optical neutrino telescopes, where individual neutrino interactions appear as sparse, detector-wide patterns of photon hits with precise timing information. A central aim of this work is to transform raw detector data into representations and models that are fast enough for low-latency and real-time applications, information-preserving enough for precision reconstruction, and robust to differences between simulation and real detector behavior.

We pursue these goals along several complementary directions. We develop ML architectures that operate directly on sparse detector data to enable trigger-level estimates of neutrino direction and energy, suitable both for online filtering and as seeds for more detailed reconstructions [1]. We also explore learning-based approaches to effectively increase detector resolution by modeling light propagation, including super-resolution techniques that improve angular reconstruction without requiring hardware changes [2]. In addition, we study representation learning for photon timing information, compressing complex timing distributions into compact, reusable features for downstream analyses [3], and self-supervised learning methods that reduce reliance on imperfect simulations and improve robustness to unmodeled detector effects [4]. Together, these efforts support fast, accurate, and scalable reconstruction for current and next-generation neutrino telescopes.

[1] Phys. Rev. D 108, 063017 
[2] Phys. Rev. D 111, L041301 
[3] JINST 20 P12010 
[4] https://arxiv.org/abs/2510.01733