Abstract -- Cosmic neutrinos provide a unique probe of the universe's most extreme environments, but their detection is extraordinarily challenging. In this seminar, I will outline how modern AI methods are transforming the field and enabling a new generation of discovery in neutrino astronomy. I will focus on methodological advances that lie at the interface of computer science and physics, illustrating how shared techniques can accelerate progress through four recent developments:

• Real-time deep learning on embedded systems: A neural-network-based trigger can double the neutrino detection rate compared to classical techniques. I will describe the associated challenge of processing continuous data streams exceeding 1 GS/s on low-power (~10 W) FPGAs and our solutions for low-latency inference in resource-constrained environments.
• End-to-end detector optimization: Using differentiable programming, we construct a fully differentiable simulation and reconstruction pipeline, enabling gradient-based optimization in a ~1000-dimensional design space for future detectors such as IceCube-Gen2. I will present the first results on a simplified detector model.
• Neural posterior estimation for event reconstruction: We employ a hybrid architecture combining Transformers with conditional normalizing flows on the sphere to infer the posterior distribution of neutrino arrival directions. This approach yields unprecedented angular resolution while ensuring statistically correct uncertainty quantification.
• Towards a foundation model for astroparticle physics: I will briefly sketch ongoing efforts toward a multi-messenger foundation model capable of jointly analyzing heterogeneous data from neutrinos, gamma rays, cosmic rays, and gravitational waves, to infer the properties of astrophysical sources.

Abstract -- Cosmic neutrinos provide a unique probe of the universe's most extreme environments, but their detection is extraordinarily challenging. In this seminar, I will outline how modern AI methods are transforming the field and enabling a new generation of discovery in neutrino astronomy. I will focus on methodological advances that lie at the interface of computer science and physics, illustrating how shared techniques can accelerate progress through four recent developments:

• Real-time deep learning on embedded systems: A neural-network-based trigger can double the neutrino detection rate compared to classical techniques. I will describe the associated challenge of processing continuous data streams exceeding 1 GS/s on low-power (~10 W) FPGAs and our solutions for low-latency inference in resource-constrained environments.
• End-to-end detector optimization: Using differentiable programming, we construct a fully differentiable simulation and reconstruction pipeline, enabling gradient-based optimization in a ~1000-dimensional design space for future detectors such as IceCube-Gen2. I will present the first results on a simplified detector model.
• Neural posterior estimation for event reconstruction: We employ a hybrid architecture combining Transformers with conditional normalizing flows on the sphere to infer the posterior distribution of neutrino arrival directions. This approach yields unprecedented angular resolution while ensuring statistically correct uncertainty quantification.
• Towards a foundation model for astroparticle physics: I will briefly sketch ongoing efforts toward a multi-messenger foundation model capable of jointly analyzing heterogeneous data from neutrinos, gamma rays, cosmic rays, and gravitational waves, to infer the properties of astrophysical sources.