Researchers at The University of New Mexico and LANL are developing a new artificial intelligence framework that could transform how scientists search for some of the rarest events in the universe by recovering faint signals that conventional methods often discard.

The project, Learning Beyond Threshold: AI-Driven Signal Extraction for Rare-Event Searches, is one of nearly 300 selected as part of the U.S. Department of Energy's Genesis Mission, which brings together DOE's world-class scientific capabilities, advanced AI, high-performance computing, and the nation's leading researchers to transform how scientific discovery is conducted and strengthen American leadership in science and technology.

The nine-month, $750,000 project is led by Physics and Astronomy Professor Dinesh Loomba, along with co-investigators including Dr. Ralph Massarczyk at LANL, and Professor Greg Taylor and Associate Professor Francis-Yan Cyr-Racine at UNM.

Detectors are instruments used to spot things like particles or light. Every detector carries background noise, often from its own electronics and sometimes from the environment, that can hide or smear the true nature of an event. Consider a photo of a person standing in front of the afternoon sun: up close and well-lit they show up clearly, but pushed into the glare their face washes out until it is barely recognizable.

Detectors face the same problem, and most deal with it by setting a cutoff and ignoring everything below it. That keeps the noise out, but it also throws away the weak, distorted, or overlapping signals that carry the information of interest. Earlier efforts have separated overlapping features, but not with high efficiency and often only with heavy computing, so the UNM/LANL team proposes replacing the cutoff with a more flexible AI framework. Feature detection of this kind is already familiar from everyday tools, like a phone that boxes every face in a group photo before the shutter is pressed.

"Our framework goes further: rather than discarding weak signals, it detects them, reconstructs them, and identifies what they are," the researchers said. "Returning to the overexposed photo, it would recover a second, shadowed face that the glare had nearly erased. Much of the machinery for image-like data already exists, so we also extend the approach to time-series data, where the signal arrives as a long stream and the interesting parts must be pulled from a constant hum of noise, much like isolating one voice singing along in a packed concert."

"Trained on known examples, the framework recovers precise event timing and catches faint signals that current methods miss, including ones at or below the usual noise level."

A second advance adapts reconstruction algorithms refined on controlled laboratory experiments, such as searches for the Migdal effect, or the complex data from astronomical observations like the Long Wavelength Array that UNM operates. Training combines simulations with real calibration and background data from experiments already running, closing the gap between clean simulations and messy reality.

The result is a framework that recovers signals normally thrown away, so detectors can pick up fainter events and pin down what they are more precisely, from imaging distant galaxies to catching particle interactions so rare they may happen only a handful of times a year.

For more information, visit The Genesis Mission.