When an oil refinery in Philadelphia exploded in June 2019, physicist Joshua Carmichael was at home in Los Alamos, New Mexico, taking care of his wife after a surgery. He’d been spending his days at Los Alamos National Laboratory developing a tool that could determine the likelihood that a given set of “signals” represented an explosion—or at least he hoped it could. He flipped open his laptop and started sucking in data sources: information from air and seismic sensors, observations from meteorological satellites, even social media reports of shaking. That afternoon, his system reported back: There was a 99% chance that the Philadelphia event had been an explosion.
Carmichael knew that, of course; he’d begun with that information. But the mishap was a test of his algorithms, to see if they would work on real-world data.
Carmichael and his colleagues describe their explosion-detection process in the August issue of Geophysical Journal International. Eventually he wants to see if a beefed-up version can use those signals to identify a nuclear explosion. “Our job is to provide decision makers with better information so that they can clearly make decisions about whether or not an explosion was recorded somewhere,” he says.
A need to merge data streams
By “somewhere,” Carmichael means pretty much anywhere: in space, in Earth’s atmosphere, on the ground, underwater, and deep underground. Although the Comprehensive Nuclear-Test-Ban Treaty places a moratorium on testing, US intelligence officials have suggested that Russia and China may have conducted low-yield tests, and the Trump administration has discussed whether the US should do the same. (Both China and the US have signed but not ratified the test-ban treaty.) Combine that with the potential for terrorist “dirty bombs,” and the need for a sensitive, speedy alert system becomes clear.
Monitoring for small-scale tests plays a moderating role in the nuclear ecosystem: To build a big bomb, you need to start small. “If you can monitor to very low yields, in a sense you’re limiting a country from building big nuclear weapons,” says Ferenc Dalnoki-Veress, a scientist-in-residence at the James Martin Center for Nonproliferation Studies.
In movies, when a nuke has detonated, it’s something out of the Bourne trilogy: A snazzy graphical interface, blinking red on a computer screen, simply alerts the intelligence analysts. In reality it’s a tedious and fraught process.
The Vienna-based Comprehensive Nuclear-Test-Ban Treaty Organization monitors the planet for signs of nuclear tests and distributes seismic, infrasound, hydroacoustic, and radionuclide data to member states. (The US intelligence community, Dalnoki-Veress notes, runs its own monitoring system under the Air Force Technical Applications Center.) To judge the nature of a suspected explosion, analysts look at the data in their specialized fields. Seismic experts look at earth-shaking information. Acoustic experts check out the sound waves in the air. Electromagnetic experts examine radio emissions. And atomic experts parse radionuclide distribution. Only together do those data sets say “Nuke!” Alone, each could imply, say, a small earthquake, a lightning storm, or a meteor streaking through the atmosphere.
The current compartmentalization is part of why Carmichael wants to make those determinations easier and faster, like in a spy movie. And, importantly, he wants to know exactly how sensitive the determiners will be to suspicious events. The ability to anticipate the smallest explosions that can reliably be detected is crucial, he says, “because then we can quantify what we expect to detect, versus miss.”
Building the model
Carmichael’s real-world data come from a multitude of disaggregated sources. Every state in the US has a dense seismic network, as do most countries, and Carmichael accesses those along with the US Geological Survey’s shake reports. For atmospheric and meteorological information, Carmichael taps resources like the University of Mississippi’s public models. In the future, Carmichael plans for his system to automatically search Twitter data aggregated by the Computational Propaganda Project at Oxford in the UK.
But obtaining the data is just the first step. The success of Carmichael’s idea hinges on advancing a century-old statistical technique to synthesize such signals and then feed the information into the algorithm to make the boom or no-boom call.
When scientists want to merge complicated data streams, they often turn to a statistical trick called Fisher’s method. It was developed in 1925 by statistician—and eugenicist—Ronald Fisher. The process combines p-values from all relevant data sets. The p-values, together and individually, indicate the likelihood that a given set of signals could coincidentally appear under the null hypothesis. A very small p-value means that the data a scientist sees are unlikely to arise simply from the day-to-day noise the scientist has quantified.
Astrophysicists use Fisher’s method to do multi-messenger studies, like those that examine both the neutrino and electromagnetic emissions from supernovae. Medical researchers use it to tell whether a vaccine’s supposed effects could just be chance. And climate scientists sometimes use it to figure out whether random, noisy climate variability—minus humans’ carbon spewage—increases ocean temperatures in the way we see.
In explosion monitoring, many things boost the noise that confounds conclusions. “There’s this total cacophony of geophysical signatures all around the globe,” says Carmichael. “It just so happens that nuclear explosions also produce those very same signatures.” Lightning striking, spark plugs helping engines idle, cell phones and radios communicating, and small seismic events all create background that masquerades as the signals Carmichael seeks. “All of his sensors are sensitive to things that are not the thing he’s trying to find,” says Brian Schrom of the Pacific Northwest National Laboratory, who works on radionuclide monitoring and was not part of the research.
To differentiate signal from noise and recognize a true deviation from the norm, Carmichael must first get a good grasp on what the noise looks like from day to day. Fusing all those normal states—rather than looking at them in a vacuum—is key. “Combining multiple measurements together allows you to really attack the noise part of the signal-to-noise threshold,” says Schrom.
And so, Carmichael’s aim is to separate the signal from the noise and look at the different signals together. But doing so using p-values, as Fisher’s method does, is problematic. P-values don’t actually tell scientists if they likely did observe what they’re looking for—in Carmichael’s case, an explosion. They simply indicate the chances that they did or didn’t observe mere noise.
As an example of this subtle distinction, Carmichael points to gravitational-wave detections. Suppose, he says, that researchers with the LIGO experiment had used Fisher’s method to reject the hypothesis that the famous waveform recorded in September 2015 was attributable to background noise. “Would that have meant that they detected a gravitational wave? No.” Instead, the gravitational-wave researchers needed to know in advance what the fingerprint of two smashed-up black holes would look like, and then use that to determine whether that fingerprint could explain the data.
For Carmichael’s explosion model, that means having to understand what the acoustic, seismic, and electromagnetic (plus radionuclide, for a nuclear detection) streams should look like for a certain level of event, fuse the data as they come in, and evaluate the probability that the event was an explosion of a given size.
Carmichael now has a model to calculate that, with the help of dozens of explosions he and his colleagues conducted at Los Alamos. They used a solid chemical explosive called Comp-B, which has about 1.3 times as much demolition power as TNT, to produce explosion energies of up to 110 megajoules. The monitoring equipment wasn’t part of a fancy system but was more homegrown: microphones placed on the ground, an off-the-shelf antenna, a buried seismometer.
Carmichael and his team found they could detect Comp-B explosions about one-sixth the size, in terms of the pressure waves they create, of the explosions that scientists would be able to pick up using an acoustic signature alone.
What lies ahead
But the model must mature before it’s ready for prime time, says seismic and geoacoustic specialist Stephen Arrowsmith of Southern Methodist University. The calculated nature of Carmichael’s explosions illustrates his model’s limitations: His paper shows the method testing itself when the researchers had prior awareness of the explosions’ specifics. “He’s basically assuming that you knew an explosion would happen,” explains Arrowsmith, who was not part of the research. “Operation-wise, you’d need to build that process where you don’t know something happened and you’re feeding all this data through.”
“He had to pick a place to start, and I think he picked a great place to start,” says Schrom. But it’s just that: a start. “He knows where his [test] pad is relative to sensors. He knows how fast signals propagate,” Schrom continues. “His work is the tip of the iceberg of a lot of follow-on work. He’s doing the simplest thing that’s interesting first.”
One of the un-simple things is the nuclear part. Electromagnetic, acoustic, and seismic waves travel at well-known speeds and in relatively simple, predictable trajectories. But radionuclides are more fickle. “They are subject to weather,” says Carmichael. They bump; they jostle; they diffuse. To incorporate their signature, Carmichael needs to include large-scale atmospheric models.
“What [Carmichael’s team is] doing by this data fusion is really, really important,” says Dalnoki-Veress. “I would be really interested in whether they could do that with radioxenon, because it’s always really hard to make a causal connection between a puff of radioactivity and a nuclear test.”
The method is also applicable beyond bombs. The background noise that Carmichael wants to toss out actually details how our planet—its tectonics, its atmosphere, and more—behaves all the time. In addition, Carmichael’s model can reveal when something besides a bomb rises above the noise, like a large earthquake or early rumblings of a volcanic eruption. Carmichael and colleagues are even trying to get funding to apply the model to elephant poaching in Africa. Strategically placed microphones could pick up a gunshot, and seismic detectors could sense elephants moving around. An algorithm could fuse the data and then send an alert to game wardens.
Putting the model to use in the nuclear sphere is more complicated. The process by which such a detection method goes from idea to operation is, says Arrowsmith, kind of ad hoc. “It needs somebody like Josh to really push on his side, and it needs engagement from the government side to then pull,” he says. “And then it needs somebody in between to take it over the valley of death.”
Maybe Carmichael can do it after his model is ready, if his model gets ready. After all, Arrowsmith points out, Carmichael works for a federally funded lab. He’s got connections and clearances. He knows what the people in power need. But governments are also full of inertia, and effecting change can be frustrating and slow. It is rarely Bourne-ish, and never boom and done.
