The IceCube Neutrino Observatory, a cubic-kilometer array of optical sensors buried in Antarctic ice, detects elusive subatomic particles called neutrinos. When a neutrino interacts with a molecule of ice, it produces secondary particles that emit blue light, which is then detected by IceCube.
Known as an “event,” its signals or light patterns are used to determine the arrival direction and deposited energy of the neutrino through reconstruction. Using a method called “likelihood optimization,” IceCube collaborators can infer the direction and the associated uncertainty of a neutrino event. However, this traditional approach can be time-consuming and often struggles to handle the high-dimensional unknowns in both event and detector properties. Such unknown properties include the precise location of all the energy depositions in the detector and the optical transmission properties of the naturally formed glacial ice. Likelihood optimizations require a correct parametrization of all these quantities; otherwise, the performance outcome suffers.
In a study submitted to the journal Machine Learning: Science and Technology, the IceCube Collaboration presents a new method that uses machine-learning (ML) tools to reconstruct the direction of neutrino events. The method outperformed traditional likelihood reconstructions for IceCube neutrino events. It is both faster and more precise, and it circumvents the “parametrization problem” of likelihood optimizations by directly learning from the training data, without ever formulating a likelihood.
“We were motivated to rethink how we do statistical inference in the age of modern artificial-intelligence (AI) tools and to see if we can improve on these traditional reconstructions both in capability and processing time,” says Thorsten Glüsenkamp, a postdoctoral researcher at Stockholm University and study lead.
Glüsenkamp and collaborators combined two modern AI technologies: 1) soft attention from transformers, which is integral to large language models like ChatGPT, and 2) normalizing flows, an efficient method to describe a probability distribution in combination with neural networks. The final trained model encodes the IceCube data and turns it into a probability distribution for the neutrino direction (the “posterior”), a process called “neural posterior estimation.”
A key challenge to overcome was to make the normalizing flow probability distribution behave nicely together with the transformer. The new study utilizes a normalizing flow called “neural spline flows,” which in its standard prescription is numerically unstable in this particular setting. The involved functions often jump around very quickly, i.e., they are not smooth. Thus, only via the incorporation of the right smoothness constraints into these neural spline flows does one obtain a numerically stable algorithm.
“The results directly carry over to a neutrino measurement where we can make more precise maps of the neutrino sky and detect fainter fluxes,” says Glüsenkamp. “Furthermore, it can be easily adapted for the IceCube Upgrade geometry or other future detector upgrades like IceCube-Gen2.”
+ info “Neural posterior estimation of the neutrino direction in IceCube using transformer-encoded normalizing flows on the sphere,” IceCube Collaboration: R. Abbasi et al. arxiv.org/abs/2604.19846