How IceCube was built and found neutrinos from deep space

5,160 light sensors frozen into the South Pole ice, a flood of background a million times larger than the signal, and the 2013 discovery of neutrinos from beyond the solar system. Play with an event display and check the statistics yourself.

Part 1 ended with an idea: freeze light sensors into a cubic kilometre of clear Antarctic ice, and watch for the faint blue Cherenkov light of particles made by neutrinos. This part is about turning that idea into a working telescope, and about the hardest step of all: proving that a handful of special events came from deep space and not from our own sky. You will see how the events look, why high energy is the key, and how the statistics were done, and you can redo the statistics with the Counting statistics helper and the fits with the curve fitter.

In short

  • IceCube has 86 strings of sensors, 5,160 sensors in all, between 1.45 and 2.45 km deep in the ice at the South Pole. It was finished in December 2010.
  • It sees thousands of muons from our own atmosphere every second and only a few cosmic neutrinos per year. Finding them took clever filters: energy, events starting inside the detector, and using the Earth as a shield.
  • In 2013 IceCube reported 28 very energetic neutrinos where about 11 were expected from the atmosphere: the first evidence of cosmic neutrinos. More data made it certain by 2014.

Building a telescope in the ice

IceCube was built in seven Antarctic summers, from the 2004–05 season to December 2010, with money mostly from the US National Science Foundation and partners in Europe and Asia. Each summer the team had about ten weeks of working weather, at an altitude of 2,800 m and temperatures around −30 °C.

The holes were melted with a hot-water drill: a 5-megawatt heating plant pumped near-boiling water down a hose, melting a hole 60 cm wide and 2.5 km deep in under two days. Then, before the water refroze, a cable with 60 sensors was lowered in. Each sensor, a Digital Optical Module (DOM), is a glass sphere about 35 cm across holding a 25 cm photomultiplier tube (a light detector that can count single photons) and its own electronics, which time each pulse of light to a few nanoseconds and send it up the cable digitally.

The layout:

  • 86 strings, mostly 125 m apart on a triangular grid, covering about a square kilometre.
  • 60 sensors per string, 17 m apart, from 1,450 m to 2,450 m deep. In all, 5,160 sensors.
  • DeepCore: eight extra strings packed closely in the clearest ice at the bottom, for lower-energy neutrinos (and the study of neutrino oscillations).
  • IceTop: 81 pairs of tanks of ice on the surface, recording the air showers made by cosmic rays.

Once frozen in, the sensors can never be reached again. They had to work for decades, and almost all of them still do.

What an event looks like

When a neutrino hits a nucleus in the ice it makes a spray of particles. What happens next depends on the neutrino’s type and the kind of collision:

  • A muon neutrino usually makes a muon, which flies straight for kilometres, glowing all the way: a track. Tracks point back to the sky within about a degree.
  • An electron neutrino, or any neutrino that just bumps a nucleus, dumps its energy in a few metres: a cascade, a nearly round ball of light. The direction is rough (often 10° to 20°), but almost all the energy stays inside the detector, so it is measured well.
  • A high-energy tau neutrino can make two cascades: one where the tau is made, and one where it decays some distance later: a double bang.

Choose an event type and watch the light spread out through one slice of the detector. Colours show the arrival time of the light at each sensor; sizes show how much light it got.

Interactive animation: IceCube event display (needs JavaScript).

Light in deep ice: the hardest part of the reconstruction

Light in ice is both absorbed and scattered. In the clearest deep ice the absorption length is well over 100 m and the effective scattering length tens of metres; both vary with depth, because the ice was laid down over 100,000 years of changing climate, including layers of volcanic dust. IceCube mapped this by flashing LEDs built into the sensors and recording what neighbours saw. The ice is even slightly birefringent, so light spreads differently along the direction of ice flow. Real reconstructions fit the expected arrival times and amounts of light at every sensor with tables built from detailed simulations; the simple straight-line fit we do below only shows the idea.

A needle under a mountain of hay

The problem is background. Cosmic rays hit the atmosphere all the time, and their showers make two kinds of nuisance:

  1. Atmospheric muons. Thousands per second are energetic enough to reach the detector from above. They are about a million times more common than neutrino events.
  2. Atmospheric neutrinos. Made in the same showers, they are true neutrinos, about a hundred thousand a year in IceCube. They are exactly the same particle as a cosmic neutrino.

Three ideas tame them:

  • Use the Earth as a shield. Muons cannot cross the Earth, so a track coming up through the detector must have been made by a neutrino. That removes the muons, but not atmospheric neutrinos.
  • Use a veto. If light first appears on the outer layer of the detector, an incoming muon is probably to blame. Events that start inside must be neutrinos. This works from every direction, including straight down.
  • Use energy. Atmospheric neutrinos come from cosmic rays whose spectrum falls steeply, and the parent pions and kaons at high energy tend to hit air before they can decay, so the atmospheric neutrino spectrum falls roughly as E^−3.7. A cosmic source should give a flatter spectrum, nearer E^−2. Above some energy the cosmic ones must win. There is one more gift: a downward-going atmospheric neutrino usually arrives together with muons from its own air shower, so the veto removes many atmospheric neutrinos too (the “self-veto”).

Slide the energy cut and the cosmic slope below and watch who wins.

Interactive chart: atmospheric against cosmic neutrino spectra (needs JavaScript).

2012 and 2013: Bert, Ernie and 26 more

In 2012 a search for extremely energetic neutrinos turned up two cascades, each depositing about 1 petaelectronvolt (10¹⁵ eV, about a thousand times the energy of a proton at the Large Hadron Collider) inside the detector. The team nicknamed them Bert and Ernie. Nothing from the atmosphere was expected at such energies in that much data.

The team then designed a search on purpose for such events: keep only events that start inside the detector (the veto) and deposit more than about 30 TeV, from data taken between May 2010 and May 2012. It found 28 events. Backgrounds (atmospheric muons that slipped past the veto, plus atmospheric neutrinos) were expected to give about 10.6. The excess had a cosmic-looking spectrum and came from all over the sky. Published in Science in November 2013, it was called evidence, at 4.1σ. With a third year of data the count rose to 37 events and the significance to 5.7σ (2014), and an independent analysis of up-going muon tracks from the northern sky, using the Earth as a shield, found the same cosmic flux in 2015. Neutrinos from beyond the solar system were real.

Try the statistics yourself in the “source or fluke?” game below: it is a search for a point source, but the logic is the same. Count what you see, compare it with what background alone would give, and ask how likely the excess is by chance.

Interactive simulation: source or fluke (needs JavaScript).

What the measured flux says

The flux is now measured with several event samples. Each comes out close to E^−2.3 to E^−2.9, depending on the energy range and channel, with a total of order 10⁻⁸ GeV cm⁻² s⁻¹ sr⁻¹ per flavour in E²·dN/dE near 100 TeV. Two striking facts follow. First, that energy flux is similar to the flux of the highest-energy cosmic rays and of the diffuse gamma-ray background, a hint that one population of sources may power all three. Second, the cosmic gamma-ray background is not bright enough to match the neutrinos at tens of TeV if the gamma rays escaped freely, which suggests many neutrino sources are hidden behind thick gas and light that absorbs gamma rays: the kind of cores of active galaxies seen in part 3. The flavour mix at Earth is close to 1:1:1, as expected after oscillations over cosmic distances, and a few double-cascade tau candidates have been found.

Checked with the tools

We redid three pieces of the IceCube story with the maths behind fibub’s tools. All three are our own simulations or recomputations; none uses IceCube’s data tables.

1. Timing a muon finds its speed and direction. We simulated a muon crossing the detector: 60 sensor hits along a kilometre of track, each with a 3 ns timing spread and a random delay from scattered light. A straight-line fit of hit time against position (the least-squares line behind the Best-fit line tool) gives the slope, and one over the slope is the muon’s speed: 0.300 m/ns, the speed of light (0.2998 m/ns) to within the fit’s uncertainty.

A simulated muon: hit time against position along the trackA simulated muon: hit time against position along the track-1e+301e+3-400-2000200400Position along the track (m)Hit time (ns)Sensor hits (simulated)Fit: speed 0.300 m/nsfibub.com
Simulated with our own numbers (60 hits, 3 ns timing spread, scattering delays) and fitted with the Scatter and best fit tool's least-squares line. Late, scattered photons pull the fit only slightly; IceCube's real reconstruction models them in full.

2. Two populations in one spectrum. We drew 200,000 simulated events from two power laws (slope −3.7 for “atmospheric”, −2.5 for “cosmic”, equal at 100 TeV), binned them, and tried to get the slopes back. Fitting straight lines to each end gives about −3.55 and −2.7: wrong, because the two populations overlap. A joint fit of both at once by Levenberg-Marquardt least squares (the method the curve fitter uses) recovers 3.77 ± 0.29 and 2.46 ± 0.10. This is why IceCube always fits the atmospheric and cosmic parts together.

Two kinds of neutrino in one spectrum (simulated)Two kinds of neutrino in one spectrum (simulated)10^-410^-210^010^210^410100100010000Neutrino energy (TeV, log scale)Events per TeV (log scale)200,000 simulated eventsLow-energy fit: slope −3.55High-energy fit: −2.73Joint fit: −3.77 and −2.46fibub.com
Our simulation of 200,000 events: atmospheric-like events with slope −3.7 and cosmic-like events with slope −2.5, set equal at 100 TeV, then binned. Straight-line fits to each end (Scatter and best fit maths) come out a little off because the two kinds overlap; a joint two-component fit by Levenberg-Marquardt least squares (the method the Curve fitter uses) gets both slopes back. It shows the method, not IceCube data.

3. How unlikely were 28 events? If only the expected background of 10.6 were present, the Poisson chance of seeing 28 or more is about 6.5 × 10⁻⁶, about 4.4σ. The paper quotes 4.1σ. The difference is honest and instructive: the background itself was uncertain (+5.0 / −3.6 events), and a fair test must allow for a larger background, which makes the excess a little less surprising.

If only background: how many events in two years?If only background: how many events in two years?05100102030Number of eventsChance (%)28 seenexpected 10.6fibub.com
Counting statistics maths: with a background of 10.6 the chance of 28 or more is 6.5e-6 (about 4.4σ). The published 4.1σ is lower because the background itself was uncertain (+5.0/−3.6 events), which a fair test must include.
Published Our check
Muon speed from hit times c = 0.2998 m/ns 0.3001 ± 0.0002 m/ns (simulation)
Slopes in a mixed spectrum input −3.7 and −2.5 joint fit −3.77 ± 0.29 and −2.46 ± 0.10
2013 excess: 28 seen, about 10.6 expected 4.1σ (with background uncertainty) 4.4σ (plain Poisson)

How we made the graphs

scripts/blog-figures/nobel-physics-2026.mjs makes all three charts. It imports fitLine() and levenbergMarquardt() from src/lib/sci/fit.ts (the least-squares code behind the fitting tools), and gammaP() from src/lib/sci/phcount.ts (the incomplete gamma function the Counting statistics helper uses for Poisson probabilities), plus normInv() from src/lib/sci/stats.ts to turn a probability into σ. The muon track and the spectrum are simulated with a fixed random seed, so the charts are the same every time. The only published numbers used are 28 events and a background of 10.6 (+5.0/−3.6) from the 2013 Science paper.

How we test the maths

tests/sci-tools/counting-statistics.mjs checks the incomplete gamma function against the exact identity P(1, x) = 1 − e^−x and the exact Poisson confidence limits against SciPy’s chi-squared quantiles. tests/sci-tools/curve-fitter.mjs checks the curve fitter against SciPy’s curve_fit: the fitted values, their uncertainties (weighted and unweighted) and the chi-square and its p-value for lines, polynomials, decays and Gaussians. These run on every change.

Check it yourself

  1. Open the Counting statistics helper and go to 4. Poisson probabilities.
  2. Enter Expected average count μ = 10.6 and Count k = 28. Read the chance of k or more: about 6.5 × 10⁻⁶.
  3. Now be cautious about the background: set μ = 15.6 (10.6 plus its upper uncertainty of 5.0). The chance of 28 or more rises to around 0.3 %, under 3σ. This is why the background error matters so much, and why the team waited for a third year.
  4. Open the curve fitter, paste pairs of (log₁₀ energy, log₁₀ counts) from any power-law spectrum, choose My own formula and enter log10(10^(a - b*x) + 10^(c - d*x)). With starting values a = 6, b = 3.5, c = 4, d = 2.2 it fits two slopes at once.
  5. Make up a “track”: in the Best-fit line tool place dots along a line of slope about 3.3 (ns per metre) and show the best-fit line. One over the slope is the speed in metres per nanosecond.

What you can use it for

  • Any counting experiment (radioactivity, rare events in a survey, clicks on a page): the Poisson tools tell you whether an excess is real and how much its background uncertainty matters.
  • Mixed populations in a histogram (two kinds of grain size, two kinds of particle): fit them together with the curve fitter rather than one end at a time.
  • Timing and speed problems: a straight-line fit of arrival time against distance gives speed and its uncertainty, whether for a muon, a seismic wave or a nerve impulse.
  • Next: which objects make these neutrinos, a flaring black hole 4 billion light-years away, a nearby galaxy hidden in dust, and our own Milky Way, in part 3.

Sources

  • Nobel Prize in Physics 2026, announced 6 October 2026: Francis Halzen, 'for decisive contributions to the IceCube Neutrino Observatory and the discovery of high-energy neutrinos of astrophysical origin'. NobelPrize.org and the Royal Swedish Academy of Sciences (kva.se).
  • IceCube Collaboration (2013), 'First observation of PeV-energy neutrinos with IceCube', Phys. Rev. Lett. 111: 021103 (the two events nicknamed Bert and Ernie).
  • IceCube Collaboration (2013), 'Evidence for high-energy extraterrestrial neutrinos at the IceCube detector', Science 342: 1242856 (28 events above about 30 TeV, background about 10.6 events, 4.1σ).
  • IceCube Collaboration (2014), 'Observation of high-energy astrophysical neutrinos in three years of IceCube data', Phys. Rev. Lett. 113: 101101 (37 events, 5.7σ).
  • IceCube Collaboration (2015), 'Evidence for astrophysical muon neutrinos from the Northern Sky with IceCube', Phys. Rev. Lett. 115: 081102.
  • IceCube Collaboration (2017), 'The IceCube Neutrino Observatory: instrumentation and online systems', JINST 12: P03012 (86 strings, 5,160 sensors, depths 1,450 to 2,450 m).
  • Physics World (2026), 'Francis Halzen wins Nobel Prize in Physics' (timeline: AMANDA 2000, IceCube complete late 2010, 28 events in 2013, 21 northern tracks above 100 TeV in 2015).

If you are an author or publisher and want a credit or a passage changed or removed, tell us and we will act quickly.

Tools in this article

Read next