Einstein probably wasn’t thinking about helping you find Starbucks when he wrote down his theories of relativity, but Google Maps wouldn’t work without them. In fact, these fundamental laws of nature—special relativity, general relativity, and quantum mechanics—are the foundation of all modern technology.
As a physicist, I’ll let you in on a little secret: Even though we use these laws almost every day, we don’t understand them very well. For instance, we can use models of gravity to send astronauts to the moon, yet have no clue how gravity works at smaller scales.
That tension between everyday use and deep ignorance is why we built the Large Hadron Collider. We study how matter behaves at the smallest scales by colliding protons at close to the speed of light and then capture the remnants with particle detectors the size of cathedrals. And then we do something astonishing: We immediately throw away 99.999% of the data. Why? Because most of the data is “bad,” and we don’t have nearly enough computing resources to keep it all.
To understand this concept, imagine that the LHC is a soccer game, and our detector is a sports photographer. Most of the gameplay is totally uninteresting, but even so, our “photographer” keeps clicking away hoping to get that million-dollar shot. Even if our photographer does catch something amazing, there is always the risk that her photo is overexposed, blurry, or poorly framed. We struggle with these same kinds of problems (just the particle detector version.)
But there’s a big difference between a photographer and our detectors: A human photographer can learn from her bad photos and adjust in real time. Our detectors cannot: they just keep on collecting data, and we have to wait months—sometimes years—before we can evaluate if the saved data is truly as “good” as it could be. (Or worse: realizing that the data we thought was “bad” might have held some valuable scientific information.)
I don’t know about you, but I really don’t like waiting and hoping for the best.
The solution I and others are pursuing is to build intelligence directly into the instrumentation: what we call smart detectors. As we take data, our smart detector will automatically notice when conditions change: when noise increases, when a region of the detector starts misbehaving, when the proton beam drifts. And in response it should be able to adjust voltages, refine calibrations, or re-balance which kinds of events it saves.
This is far from a novel idea: the camera on my phone automatically adjusts to ensure that I’m always getting the best possible pics. So why can’t our particle detectors?
This brings us to the crux of the problem: We are not just toggling the exposure on a camera up and down. We are talking about letting algorithms influence voltages and calibrations on delicate, expensive hardware, and shape which data we keep and which we discard. We worry that a runaway feedback loop or a misjudged pattern could cause us to miss a Nobel Prize–worthy discovery or even damage parts of the detector.
That is why we are proceeding cautiously. We are adapting commercially available reinforcement learning methods and testing them inside digital twins of our experiments. From this, we can compare the smart detector performance with our more traditional methods for data collection and evaluate the risks versus the benefits. Our near-term goal is to build a physical prototype to see if this can be done safely and effectively before anything is deployed more broadly.
If we succeed, the benefits are not abstract. It would shorten the time between taking data and understanding what that data really means. It would free students and researchers from spending their nights hand-tuning voltages and chasing down noise, so they can focus on the physics itself. Most importantly, it could compress the timelines of discovery. In the past, it took decades for ideas like relativity or antimatter to turn into practical tools like GPS and PET scans. By building real-time learning directly into our instruments, we may be able to accelerate the time between paradigm shifting discovers from a human lifetime to a single PhD thesis.
