Somewhere beneath the earth, a vat of supercooled argon holds a single flash of light and motion that may carry a whisper from the sun, the stars, or the first moments of the universe. Reading that flash is the work of SPINE, a machine-learning program from SLAC that now helps interpret these crowded pictures across roughly two dozen institutions. Its early trial on real data from the ICARUS detector found no sign of a hypothetical fourth neutrino, yet proved that the method can be trusted beyond simulation. As the field prepares for DUNE, the question is no longer whether such tools can see c
SLAC's SPINE AI Tool Decodes Ghost-Particle Data for Neutrino Experiments Worldwide
"It could be the difference between seeing something and not seeing anything."
Mark
Why does a reconstruction tool get this much attention? Isn't the detector the real instrument?
Mimi
Because a detector only records pixels. SPINE turns them into particles and interactions. Drielsma says a working reconstruction separates an experiment that can measure from one that can't.
Mark
What's different about how it works?
Mimi
Older approaches chained single-task algorithms, so one early mistake carried through. SPINE uses layered models: features, then particles, then whole interactions. It shows the particles behind its answer.
Luke
That transparency is the developers' description, though. What's actually demonstrated?
Mimi
The ICARUS analysis. SPINE recovered roughly 60% more usable interactions than conventional reconstruction, with a cleaner sample and reduced systematic uncertainties.
Luke
Worth saying that's one analysis, and the 60% figure comes from the SLAC team's own account. And no sterile neutrino signal was found.
Mark
So the result is about the tool, not the discovery?
Mimi
Yes. It showed SPINE can be trusted on real data, not just simulation.
Luke
DUNE's near detector is the open question. The team says SPINE is the only approach that can handle it, but DUNE hasn't taken data yet, and that claim is untested at scale.
Mimi
Right. Two-detector Short-Baseline results come first, expected in 2026 and 2027.
The Pulse
- A single liquid argon image can hold many neutrino interactions tangled together with cosmic-ray noise, making every reading a race against confusion.
- Conventional chains of single-task algorithms let early mistakes ripple through the entire reconstruction, so researchers needed a more resilient approach.
- SPINE's layered models recovered about 60% more usable interactions in the first ICARUS analysis, with cleaner samples and reduced uncertainties.
- The ICARUS sterile neutrino search found no signal, but the result proved the tool performs on real data, and DUNE's near detector will soon test it against roughly 1,000 pile-up events every 10 seconds.
Somewhere beneath the earth, a vat of supercooled argon holds a single flash of light and motion that may carry a whisper from the sun, the stars, or the first moments of the universe. Reading that flash is the work of SPINE, a machine-learning program from SLAC that now helps interpret these crowded pictures across roughly two dozen institutions. Its early trial on real data from the ICARUS detector found no sign of a hypothetical fourth neutrino, yet proved that the method can be trusted beyond simulation. As the field prepares for DUNE, the question is no longer whether such tools can see clearly, but whether they can keep their footing when the detector is flooded with overlapping events.
Inside a liquid argon detector, a single image can hold many neutrino interactions at once, each one trailing particle tracks and showers, all mixed up with cosmic rays and background noise. Sorting that picture out is the job of SPINE, a machine-learning program developed at SLAC National Accelerator Laboratory that now runs on eight detectors at about 25 institutions worldwide.
Neutrinos are among the most abundant particles in the universe, yet they pass through people by the trillion every second without a trace. Studying them might reveal clues about dark matter and why matter exists at all, which is why scientists build enormous detectors and intense beams. Reconstructing each rare interaction, however, has been its own challenge. Older methods chained together specialized algorithms, so an early error could spread through everything downstream. SPINE works in stages instead, first spotting basic features, then identifying individual particles, then assembling the whole event. Because it reconstructs each particle, its reasoning can be inspected rather than hidden inside a black box.
Its first real test came with ICARUS, the largest detector in the Short-Baseline program, which searched for a hypothetical sterile neutrino. The analysis found no sterile signal, but it showed SPINE could produce trustworthy results from real data, recovering roughly 60% more usable interactions than conventional methods. Drielsma summed up the gap between simulation and reality as the difference between things working and things being trusted.
The stakes rise with DUNE, which will send a beam 800 miles from Fermilab to underground detectors in South Dakota to study how neutrinos change flavor and whether they behave differently from antineutrinos, a difference that could help explain why the universe contains more matter than antimatter. Its near detector will see about 1,000 overlapping interactions every 10 seconds, and SPINE is currently the only approach trained to handle that crowding. Combined ICARUS and Short-Baseline Near Detector results are expected in 2026 and 2027, and then comes DUNE. As Drielsma put it, "Until a year ago, we hoped this approach would work. Today, we believe it will."
A single image from a liquid argon detector can hold many neutrino interactions at once, each throwing off several particle tracks and electromagnetic showers, all tangled up with cosmic rays and other background activity. Sorting out what happened in that picture is the job of SPINE, a machine-learning program built at the Department of Energy's SLAC National Accelerator Laboratory. It now runs on eight detectors at roughly 25 institutions around the world.
The difficulty starts with the particles themselves. Neutrinos are among the most abundant particles in the universe, produced by nuclear reactions in the sun, exploding stars, accelerators and radioactive decay on Earth. They are nearly massless and interact so weakly with matter that trillions pass through each person every second without a trace. Studying them could bear on questions from the nature of dark matter to why matter exists at all. So scientists build huge detectors and intense beams to catch a few interactions, and then face the harder task of reconstructing each one.
SPINE, short for Scalable Particle Imaging with Neural Embeddings, works on data from liquid argon time-projection chambers. These hold hundreds to thousands of tons of ultra-pure argon kept near -300 F. When a neutrino strikes an argon atom, the resulting particles leave trails of electrons that drift to collection wires and build a three-dimensional picture. The technology is used by the Short-Baseline Neutrino program at Fermilab in Illinois and by the Deep Underground Neutrino Experiment, DUNE, which is under construction in Illinois and South Dakota and expected to start taking data in 2030.
For years, researchers handled this by chaining specialized algorithms, each doing one narrow task and passing results along, so an error early on could ripple through everything after it. SPINE instead uses several machine-learning models that work in stages: spotting basic features, identifying individual particles, then assembling the full interaction. "We built the reconstruction hierarchically," said François Drielsma, the lead developer. Because it reconstructs the individual particles behind a signal, it shows its work rather than acting as a black box. Kazuhiro Terao started the project, applying computer-vision techniques to neutrino images in 2015 and 2016 as a postdoctoral researcher with a DOE Early Career Award. Drielsma, then a postdoc, helped turn it into SPINE.
Its first test on real data came with ICARUS, the largest detector in the Short-Baseline program, which is searching for sterile neutrinos, a hypothetical fourth type beyond electron, muon and tau. Sterile neutrinos would add oscillation patterns at short distances that the three known flavors cannot explain. The analysis looked for muon neutrinos disappearing at an unexpected rate. It found no evidence of a sterile signal, but it showed SPINE could produce trusted results from real data and not just simulation. Compared with conventional reconstruction, SPINE recovered roughly 60% more usable interactions, with a cleaner sample, better energy resolution and reduced systematic uncertainties. Drielsma noted that making things work in simulation and in data "to the degree that we trust the results" are very different things.
DUNE raises the stakes. It will send a beam from Fermilab to underground detectors in South Dakota, 800 miles away, to measure how neutrinos change flavor, determine the ordering of their masses and look for differences between neutrinos and antineutrinos. Such differences could help explain why the universe holds more matter than antimatter. The far detector will record relatively few interactions, but the near detector will see about 1,000 pile up every 10 seconds, according to Drielsma. The SLAC team has trained SPINE models for both detectors, and for the near detector it is currently the only reconstruction approach able to handle that overlap.
SLAC physicist Mark Convery described SPINE as the third leg of a stool alongside an intense beam and a capable detector, because the effects now being sought are subtle. Terao called it "a very rare beast" that does many tasks at once.
The next test is a combined analysis of ICARUS and the nearby Short-Baseline Near Detector, comparing observations at two distances from the beam, with results expected in 2026 and 2027. After that comes DUNE. "Until a year ago, we hoped this approach would work," Drielsma said. "Today, we believe it will."
Notable Quotes
Until a year ago, we hoped this approach would work. Today, we believe it will. We've built the tool. Now we get to see what it discovers.— François Drielsma, SPINE lead developer, SLAC
SPINE is the third leg of the stool, alongside an intense beam and a capable detector.— Mark Convery, SLAC physicist (paraphrased)
Want the full story?
Read the original at HPCwire ↗
Science & Space