IceCube Competition on Kaggle

In early 2023, the IceCube Neutrino Observatory sponsored a competition on Kaggle called IceCube – Neutrinos in Deep Ice. As a former physics and astronomy student, and an on-and-off Kaggle participant, I was intrigued.

Neutrino Physics and Astronomy

Neutrinos are electrically neutral particles given off by beta decay, as well as by high energy events such as supernovas and gamma-ray bursts. Neutrinos interact only very weakly with other particles. As the IceCube observatory FAQ notes, 100 trillion neutrinos pass through your body each second, but one would interact with you only about once per 100 years. Therefore, neutrino detectors rely on interactions in large volumes of water or ice.

While the weak neutrino interactions make it hard to detect neutrinos, they also means that neutrinos can travel directly from astronomical sources as near as our Sun, and as far as other galaxies, whereas light can be absorbed and scattered by intervening material. In 1968, Davis, Harmer, and Hoffman reported the detection of electron neutrinos from the sun. Their detector use a 100,000 gallon tank of dry-cleaning fluid buried 1478m deep in the Homestake Gold Mine in South Dakota. This discovery confirmed experimentally that nuclear fusion powers the sun, created the field of neutrino astronomy, and led to 2002 Nobel Prize in Physics for Davis and Koshiba.

The Homestake experiment observed only 1/3 the expected number of solar neutrinos (as previously calculated by John Bahcall, see Bahcall, Bahcall, and Shaviv 1968). Physicists and astronomers debated the cause of this deficit, known as the “solar neutrino problem”, for decades. Ultimately, they concluded that neutrinos oscillate among 3 “flavors”, while Davis’s experiment was sensitive only to one flavor, the electron neutrino. Later measurements of both electron neutrinos and the total neutrino flux confirmed this theory, which also implied that neutrinos have mass, leading to a 2015 Nobel Prize in Physics for Kajita and MacDonald.

Experimentally, neutrinos are detected indirectly. The original Davis experiments measured argon produced by reactions between neutrinos and chlorine. Similarly, IceCube detects light known as Cherenkov radiation, given off by high-energy charged particles passing through ice faster than the speed of light in ice.¹

Machine Learning or Physics?

The goal of the Kaggle contest was to find a fast and accurate way to find the direction of the original neutrino. One of the ways that IceCube distinguishes astronomical neutrinos from particles created by cosmic rays hitting the Earth’s atmosphere is by limiting analysis to particle tracks traveling up through the ice, which implies that they have come through the bulk of the Earth’s mass. Furthermore, being able to quickly identify the source of a neutrino event allows astronomers to obtain information about the source from optical telescopes pointed in the same direction as the neutrino event.

This reconstruction task was framed as a machine learning challenge; the IceCube observatory published a large training set of event data (position, timing, and energy of detections) together with directions, and competitors were asked to write software which could reconstruct the direction from event data alone.

From the start, I wondered whether my physics background would give me a leg up. So rather than starting with machine learning, I started thinking about the physics.

To learn more, continue to Part 2 of this article.

1. The theory of special relativity forbids matter from traveling faster than the speed of light in vacuum known as $c$. However, in matter with a index of refraction $n > 1$, wavefronts of light travel only at $c/n$. In Cherenkov radiation, the particle sends out light waves in a cone, similar to how an object traveling faster than sound sends out sound (pressure) waves creating a sonic boom ↩︎