New machine learning method dramatically improves IceCube data processing

Machine learning has arrived at the South Pole.

Well, not literally, but machine learning is now being applied to data collected at the South Pole by the IceCube Neutrino Observatory, an unconventional telescope made up of thousands of sensors buried in ice. IceCube’s goal is to detect tiny, nearly massless particles called neutrinos that fly through space at high energies and find out where they came from.

But IceCube sees much more than high-energy neutrinos; its sensors observe around 2,600 signals per second. To separate the wheat from the chaff, IceCube must analyze its data in real time to see which events are interesting through a process called “reconstruction.” It’s a time- and computer-intensive process, but IceCube researchers recently found a way to improve it with an exciting tool in machine learning called convolutional neural networks (CNN).

In a technical paper recently published in JINST, the IceCube Collaboration describes its first convolutional neural network reconstruction method. Compared to IceCube’s standard methods, the CNN method improves accuracy of the reconstruction and reduces the time necessary to run the reconstruction by two to three orders of magnitude. It has already been used for one IceCube analysis, and it is currently being used for several other IceCube projects.

This plot shows the angular resolution for cascade events as a function of energy. The y-axis shows the difference between the reconstructed and true directions, where lower values mean that IceCube’s reconstructed direction is closer to the true direction (obtained from Monte Carlo simulations). The new reconstruction method is represented by blue lines, and the standard reconstruction method is represented by gray. The CNN starts to outperform the standard reconstruction at around 104 GeV; events above this are interesting because they are more likely to be neutrinos from extraterrestrial sources. Credit: IceCube Collaboration

When IceCube makes a detection, reconstruction starts immediately with computers on site at the South Pole, but hardware limitations mean each event must be processed quickly to prevent pileup. Even though there are sophisticated and powerful reconstruction methods, they can take minutes—even hours—to reconstruct a single event, which is too long for the South Pole system and too computationally complex for analyses done elsewhere.

So IceCube collaborators sought a robust reconstruction method that could handle raw data in a short amount of time. Ideally, the new method would also reduce uncertainties on reconstruction parameters compared to the current methods. They turned to machine learning.

A branch of artificial intelligence, machine learning is the science of training computers to improve automatically, without explicit instruction, by analyzing and finding patterns in data. Here in 2021, machine learning is used in many areas of daily life, from the recommendations curated by your streaming services to the algorithms that arrange your social media timeline. More recently, it has begun to establish itself as a tool for doing science.

For this new reconstruction method, IceCube collaborator Mirco Huennefeld, a PhD student at Technical University Dortmund in Germany, used a specific type of machine learning called deep learning that works by studying “neural networks” with many layers of artificial neurons—sets of algorithms arranged like neurons in a biological brain. More specifically, Huennefeld decided to employ convolutional neural networks (CNNs), which are neural networks used primarily to classify images. CNNs are useful because they take advantage of translational invariance (it doesn’t matter where in the image a feature is) and the importance of local pixels to simplify the task computationally. Unlike image recognition with traditional machine learning that looks at millions of pixels all at once, a CNN can focus on a small part of an image and try to identify simple features there.

“Image data and events recorded by IceCube share similar underlying patterns such as the translational invariance and the importance of local pixels/DOMs,” said Huennefeld, a lead on the new paper. “CNNs can exploit this information and therefore seemed like a viable approach.”

The new reconstruction method proved to be orders of magnitude faster than IceCube’s previous methods. It was especially effective for a specific type of event called cascades. In fact, the method was used in a 7-year cascade analysis published last year and was recently implemented in a new alert system that notifies collaborators of cascade detections in real time (it has already sent out a couple of alerts). Interestingly, in addition to being faster, the new method also significantly improved the angular resolution at higher energies—by as much as 50 percent.

It’s only the beginning for deep learning applications in IceCube; more exciting things are on the horizon. “This paper is a big step for IceCube toward the adoption of deep-learning–based reconstruction methods,” said Huennefeld. “I believe this is the first IceCube Collaboration journal publication on deep learning. I assume it will be somewhat of a door opener for the many projects that are already underway.”

+ info “A Convolutional Neural Network based Cascade Reconstruction for the IceCube Neutrino Observatory,” IceCube Collaboration: R. Abbasi et al., JINST 16 P07041 (2021),,