Dark Matter Detection Is Outrunning Moore's Law, and All It Got Was One Event
The one nuclear recoil event LZ bought with seven tonnes of liquid xenon clears 3σ locally, but once you fold in the look-elsewhere effect it's down to 2.6σ — which is why the team has to keep saying it might just be a background fluctuation.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
AI cannot conjure an event out of the analysis chain
LZ uses AI heavily in its analysis chain, but the constraints are far stricter than in ordinary settings: any step that uses AI must be rigorously understood as to why it works that way. The reason is straightforward — when you are discussing a single event, they absolutely cannot accept the analysis chain popping the event out because some AI component "magically" did so, and then have the team vouch for that result. This treats interpretability as a precondition for publication, not an explanation bolted on afterward.
— Rick GateskillStudents thought their advisor was assigning them problems with no answer
Rick describes a feedback session with his graduate students: one student wrote in their feedback that they felt the advisor was deliberately assigning them work that would never produce an answer, because the research kept producing negative results. His conclusion was that the team had failed completely — they had not made clear to the student that this is simply how research works. His figure: more than 95% of the time you get no positive result, and even good research, well motivated and imaginative, Nature will still ignore entirely.
— Rick GateskillA dark matter particle crossing ten light-years of lead
How weak the interaction between dark matter and ordinary matter is: shoot a hypothetical dark matter particle into lead, and stack the lead all the way out past the nearest star beyond the Sun and then more than twice again — that is, 10 light-years — and the probability that it interacts while crossing that much lead is still less than half. But the flux is enormous — roughly a hundred million pass through your body every second. So as long as the detector mass is large enough, over weeks to months the probability of a single interaction starts to become finite rather than zero.
— Rick GateskillDark matter detection is outrunning Moore's Law
Over 40 years, underground experiments have covered nearly eight orders of magnitude in the dark matter–nucleon scattering cross section. Why such a huge span is possible: if the exchanged propagator has mass, in the low-momentum limit the propagator typically carries a 1/mass⁴, so a one-order-of-magnitude difference in the propagator can mean four orders of magnitude in the cross section. Rick says they were forced to demand of themselves in turn that the pace of improvement in detector scale, sensitivity and background rejection keep up. Moore's Law is about one order of magnitude every 10 years; they do about one order of magnitude every six years.
— Rick GateskillA ten-gram detector achieved the sensitivity of seven tonnes
The dark matter detector Rick used during his PhD was only about 10 grams, while LZ today is seven tonnes of liquid xenon, yet the sensitivities are basically comparable — in between is a scale change of more than a million times. That leap came from measuring single electrons and single photons: in liquid xenon, a particle interaction produces scintillation light and ionisation electrons, the electrons are drifted by an electric field to the liquid surface and read out, and these single-photon, single-electron-level measurements are combined to determine that an interaction really did occur at the center of the detector that ordinary background or radioactivity cannot account for.
— Rick Gateskill220 days of data yielded exactly one event
After unblinding 220 effective days of data, only one event remained, close to the nuclear recoil band but sitting high. The low-energy event was an accidental coincidence; in the high-energy region accidental coincidences are not the dominant background. The main background contribution in the high-energy region is high-energy gamma rays that may scatter multiple times inside the detector. To estimate this kind of background, the team simulated the equivalent of 220,000 days of data — because they had to rule out the extremely rare misidentification of multi-site single-ionisation (MSSI), whose rate is on the order of one in a million.
— SpeakerOne event fell from 3.4σ to 2.6σ
The nuclear recoil event LZ saw had a local significance above 3σ, but once the look-elsewhere effect is included — because the signal could appear over a very broad range of recoil energies, and there are many physical models that could produce a dark matter recoil — the global significance drops to about 2.6σ, that is, a probability of one or two percent. Rick says this is why they have to keep saying "this could be a background fluctuation." He also says there is no unique correct way to statistically combine a pile of possible models and background contributions, and that part of the reason they published the result was to hear from colleagues which combination method feels most natural.
— Rick GaitskellThe neutrino fog: 20 boron-8 neutrino events
Last December LZ announced it had accumulated about 20 neutrino events, from the 8-boron component of solar neutrinos, at 15 MeV, which can leave an observable nuclear recoil in xenon through coherent scattering. This means that when searching for dark matter at the lowest energies, dark matter events will be mixed together with low-energy boron-8 neutrino events. Going to higher energies, solar neutrinos cut off at 15 MeV, but atmospheric neutrinos have much higher energies. Rick says that if you go from 10 tonnes to 100 tonnes and run for ten years, you will start to see nuclear recoils caused by atmospheric neutrinos and even the diffuse supernova neutrino background — this is "entering the fog": you have to measure the neutrino signal well enough to see whether there is an anomaly above the rate.
— Rick GaitskellIn their own words · checked verbatim
we would not be in a situation where an analysis chain simply popped an event out of magic because of an AI component. And that was what we stood behind.
Rick Gateskill5:28
So we just realized we'd utterly failed our students in terms of really telling them about how research works, which is most of the time.
Rick Gateskill8:54
if we fired a single dark matter particle, hypothetical dark matter particle, through lead, we could actually pile that lead all the way out to the closest star beyond the sun … and actually go about twice, more than twice that distance, so 10 light years. And even though this dark matter particle was traveling through lead, less than 50-50 chance it would have interacted at that point.
Rick Gateskill16:16
It's just saying that the drift time of the electron and the location of where the electron hits the surface and lights up like a Christmas tree, that's the projection bit.
Speaker45:15
the first rule of physics is not to fool yourself, and the second rule is you're the easiest person to fool.
Speaker1:21:11
adjectives don't cut it you have to associate quantitative numbers with things
Rick Gaitskell1:26:22
we do so many experiments that you have to get unlucky in terms of a background fluctuation, which is why we often talk about five sigma as being a necessary requirement if you're really going to say a signal is robust enough to claim direct discovery
Rick Gaitskell1:35:38
If an experiment or particular technology wants to work, I call that a low Sisyphean Index.
Rick Gaitskell1:57:23
Figures
| Number of authors on the LZ paper | More than 250 | 5:28 |
| Dark matter share of the universe's composition | About 25% | 13:04 |
| Ordinary baryonic matter share of the universe's composition | Less than 5% | 13:04 |
| Mass of LZ's liquid xenon active volume | About 7 tonnes | 37:58 |
| Dimensions of the liquid xenon detector | Diameter and drift length each about 1.5 meters (150 centimeters) | 42:07 |
| Track length of electrons read out in the gas | About the last 20 millimeters | 45:15 |
| VUV wavelength | About 175 nanometers | 50:30 |
| PMT array | 250, spanning about 1.5 meters | 51:40 |
| Local significance of the event | More than 3 sigma | 1:25:18 |
| Global significance of the event | About 2.6 sigma, about one or two percent | 1:25:18 |
| Number of neutrino events accumulated by LZ | About 20 (8-boron solar neutrinos, 15 MeV) | 1:38:48 |
Glossary
- look-elsewhere effect
- The signal could appear over a very broad energy range, so the local significance must be converted into a global significance according to the search range.
- Sisyphean Index
- A term Rick coined: how much the technology wants to work on its own; the lower the index, the easier it is to push forward.
- MSSI
- A high-energy gamma scattering multiple times inside the detector while producing only one ionisation, which can be misidentified as a dark matter event.
- TPC
- A detector that reconstructs the three-dimensional position of an interaction from electron drift time and the position of light on the liquid surface.
How to listen
Founders and investors in hard tech and deep tech, and engineers who care about the limits of AI interpretability — watch how a real experiment refuses to let an AI black box vouch for its own result.
The DAMA and MOND discussion from 01:42:50 to 01:47:58, which is loosely connected to the main mechanism.