Artificial Intelligence · Materials Science
An experiment that used to take months for answers.
Now it talks back before the beam even switches off.
Argonne National Laboratory has built an AI tool called DONUT that reads complex X-ray data in real time, something that used to take scientists weeks or months of manual analysis. By baking real physics directly into the neural network, DONUT skips the usual need for labeled training data entirely — and it's already reshaping how experiments run at one of the world's brightest X-ray facilities.
The Advanced Photon Source: a facility that fires focused X-ray beams at materials to reveal their structure, atom by atom.
Signal Depth Navigator
Researchers at Argonne National Laboratory built DONUT, a physics-aware neural network that interprets X-ray nanodiffraction data in real time, without needing labeled training examples. This article covers why this kind of X-ray data has always been so hard to analyze, how baking physics into the AI changes that, what real-time results actually unlock for scientists, and how this fits into a much larger national push to accelerate scientific discovery with AI.
Understand
2. What Scientists Are Actually Trying to See 3. Why This Data Has Always Been So Hard to Read 4. Teaching AI Physics, Not Just PatternsEvaluate
5. No Labels Needed — Why That Actually Matters 6. A Recipe Rewritten for Every New Question 7. What Changes When Answers Arrive InstantlyLook Ahead
8. Experiments That Choose Their Own Next Step 9. Opening the Door to New Researchers 10. A Piece of a Much Bigger National PushFinal Take
11. Manohar InsightAn Experiment That Used to Take Months Now Takes Seconds
Picture running a demanding, expensive experiment, then waiting weeks just to find out if it actually worked.
That's been the reality for scientists studying advanced materials at the atomic scale, for years.
Researchers at the U.S. Department of Energy's Argonne National Laboratory just changed that timeline dramatically.
Their tool is called DONUT, and it delivers results while the experiment is still running, not weeks after it ends.
It operates at the Advanced Photon Source, or APS, a DOE Office of Science user facility that fires extraordinarily bright, focused X-ray beams at materials.
The work was published in the journal npj Computational Materials.
Aileen Luo, an assistant computational scientist at Argonne and Cornell University, put the shift plainly.
"Instead of waiting for weeks to find out if an experiment worked, we can now get answers on the spot," she said.
That single change, answers on the spot rather than weeks later, is the entire story here.
What Scientists Are Actually Trying to See
To understand why DONUT matters, it helps to know what technique it's actually speeding up.
That technique is scanning X-ray nanodiffraction microscopy, or SXDM for short.
SXDM works by scanning a tightly focused X-ray beam across a sample, point by point, collecting information about its crystal structure.
That structural information reveals how materials actually behave, down at the nanoscale — a level of detail invisible to almost any other technique.
This isn't a niche academic curiosity. It matters directly for how well batteries perform, how effectively catalysts speed up chemical reactions, and how reliably advanced electronic or magnetic devices function.
All of those technologies depend on structure at scales far too small to see directly, even under a powerful microscope.
SXDM is genuinely one of the more powerful tools scientists have for actually seeing that structure.
The catch has never been collecting the data. Modern X-ray facilities are extremely good at that part.
The catch has always been what comes next: actually making sense of what the beam brought back.
Why This Data Has Always Been So Hard to Read
SXDM produces data with multiple layers and dimensions, all packed together, all needing careful interpretation.
That complexity is exactly what has made analysis so genuinely difficult for scientists, for as long as the technique has existed.
Collecting X-ray diffraction data has never been the hard part — making sense of it, layer by complex layer, has been.
The traditional approach relies on manual comparison. Researchers compare the measured X-ray patterns against simulated ones, hunting for a match.
That process is genuinely time-consuming. It's also prone to error, since it leans heavily on careful, painstaking human judgment.
Multiply that effort across an entire scan, covering thousands of individual points on a sample, and the workload becomes enormous.
That's exactly why analysis has traditionally taken weeks or months, even after the actual X-ray measurement itself was finished in hours.
The bottleneck was never the X-rays. It was always the humans trying to keep up with what the X-rays revealed.
Solving that bottleneck meant building something that could interpret this complex, layered data automatically, and fast.
Teaching AI Physics, Not Just Patterns
DONUT stands for Diffraction with Optics for Nanobeam by Unsupervised Training.
That name hides a genuinely important design choice: DONUT is a physics-aware neural network.
Most AI models learn purely from examples, spotting statistical patterns without any built-in understanding of the underlying science.
DONUT works differently. It's built with an actual understanding of the physical laws governing how focused X-ray beams interact with materials.
In other words, it doesn't just recognize what a diffraction pattern tends to look like. It understands, structurally, why that pattern forms in the first place.
The tool was developed and tested using data from the Hard X-ray Nanoprobe beamline, jointly shared by the APS and Argonne's Center for Nanoscale Materials, or CNM.
The CNM is itself a separate DOE Office of Science user facility, based at Argonne.
That real-world beamline data gave DONUT exactly the grounding it needed, tying the physics model directly to genuine experimental conditions, not idealized simulations alone.
No Labels Needed — Why That Actually Matters
Here's the detail that makes DONUT genuinely different from most machine learning tools in this space.
Most AI systems need labeled training data. Each example has to be matched with the correct answer beforehand, usually by an expert, or through separate simulation work.
Preparing that labeled data is itself a slow, specialized task. It demands real expertise, and it eats up time before the actual analysis can even begin.
DONUT sidesteps that requirement almost entirely. It learns directly from experimental data, without needing pre-labeled examples at all.
That's what "unsupervised training," right there in its own name, actually refers to.
For busy beamline users, working against tight scheduled experiment time, that's a genuinely major advantage.
No pause to prepare a labeled dataset. No dependency on an outside expert's schedule, just to get the analysis pipeline started.
The physics baked into the model does the heavy lifting that labeled examples would otherwise have to provide.
That's a genuinely elegant substitution — real physical understanding, standing in for a mountain of manually prepared training data.
A Recipe Rewritten for Every New Question
Beyond speed, DONUT offers something else scientists genuinely value: flexibility.
Mathew Cherukara, a computational scientist and group leader at Argonne, describes exactly how adaptable the tool is.
"DONUT is flexible and customizable," he said. "You can train it on the data you collect at the start of the experiment and even adjust what you want it to predict during the experiment."
That's a genuinely different workflow from most rigid, pre-built analysis tools.
A scientist doesn't need to commit, upfront, to exactly what question the analysis will answer.
As the experiment unfolds, and interesting new questions emerge, DONUT can shift focus right along with them.
Cherukara frames this playfully, but the underlying point is genuinely serious: "It's like having a fresh DONUT recipe for every new scientific question."
Science rarely follows a perfectly linear path. Interesting results often raise new questions nobody anticipated at the start.
A tool that can bend with that unpredictability, rather than forcing researchers back to square one, is a genuinely practical kind of flexibility.
What Changes When Answers Arrive Instantly
It's worth pausing on just how much faster this actually is, in concrete terms.
DONUT delivers results sometimes hundreds of times faster than traditional analysis methods.
That's not a modest efficiency gain. It's a fundamentally different research rhythm.
Luo captures exactly what that speed unlocks: "DONUT lets us see what's happening inside materials as the experiment unfolds."
"That means more productive experiments and more opportunities for discovery," she added.
Consider what real-time feedback actually changes in practice. A researcher can spot a problem mid-experiment and fix it immediately, rather than discovering it weeks later, once beam time is long gone.
A researcher can also spot something unexpectedly interesting, and immediately decide to probe it further, right then and there.
Neither option exists when results only arrive weeks after the beam has already switched off.
Luo puts it simply: "No more waiting for the dough to rise." The experiment and the understanding of it, finally, happen together.
Experiments That Choose Their Own Next Step
Real-time analysis doesn't just save time. It opens up entirely new kinds of experiments.
Real-time interpretation is what makes it possible for an experiment to react to its own results, rather than waiting for a human to catch up.
DONUT's speed and accuracy make so-called "self-driving" research genuinely possible at the APS.
In autonomous experiments like these, the next measurement gets chosen automatically, based directly on the latest results, without waiting on a human to review and decide.
That capability matters enormously for studies testing materials under real-world conditions, where circumstances can shift quickly and unpredictably.
Battery materials cycling under stress, or catalysts reacting in real time, are exactly the kind of dynamic systems where waiting weeks for analysis simply isn't an option.
If a material starts behaving unexpectedly mid-experiment, researchers need to respond right away, not after the fact.
Autonomous, self-adjusting experiments are exactly how that responsiveness becomes possible at scale.
This is where DONUT stops being simply a faster analysis tool, and starts becoming part of the experimental process itself.
Opening the Door to New Researchers
There's a quieter benefit tucked inside this story, easy to overlook next to the flashier "real-time" headline.
DONUT genuinely lowers the barrier to entry for new users at both the APS and the CNM.
That includes graduate students and visiting scientists — researchers who may not yet have deep, specialized expertise in X-ray data analysis.
Traditionally, preparing labeled datasets required real specialized knowledge and a meaningful time investment before any analysis could even begin.
That requirement has historically limited who could realistically make full use of these powerful facilities.
Eliminate the need for expert-labeled data, and that barrier substantially drops.
A newer researcher can arrive at the beamline, run their experiment, and start genuinely interpreting results, without first needing years of specialized analysis training.
That's not just a convenience. It's a meaningful shift in who gets to participate fully in cutting-edge materials research.
A Piece of a Much Bigger National Push
DONUT's usefulness is arriving at a particularly well-timed moment.
The APS itself has recently been upgraded, now delivering brighter X-ray beams and collecting data at significantly higher speeds than before.
Brighter, faster data collection is only genuinely useful if analysis can actually keep pace with it. DONUT is exactly what makes that pace sustainable.
The team isn't stopping at current capabilities, either. Work is already underway on new versions of DONUT built for autonomous microscopy specifically.
The goal there: AI tools that can help guide portions of an experiment, without requiring constant human input at every step.
Researchers are also exploring whether DONUT's physics-aware approach could extend to other advanced imaging techniques, ones that generate similarly complex, multi-layered datasets.
And there's an even bigger picture here. DONUT's physics-aware training framework is expected to support the DOE's Genesis Mission.
Genesis is a bold national initiative aiming to double scientific productivity and accelerate innovation through AI, and DONUT forms the foundation for its light and neutron source project specifically.
A single laboratory tool, quietly becoming a building block for national-scale scientific strategy.
Manohar Insight
What I find genuinely compelling here isn't the speed number, impressive as it is.
It's the decision to teach the AI actual physics, rather than simply feeding it more data and hoping patterns would emerge on their own.
That's a meaningfully different philosophy from a lot of AI research right now, where scale often gets treated as the default solution to everything.
Here, understanding replaced brute force. The model didn't need mountains of labeled examples, because it already understood, structurally, what it was looking at.
There's something quietly reassuring in that approach, especially for scientific tools specifically. Physics doesn't change from one dataset to the next. Grounding an AI in it seems like the more durable bet.
I also can't help noticing how much of this story is really about access. Lower barriers for new users. Faster answers for busy beamline schedules. A foundation for a national initiative meant to accelerate discovery broadly, not just for a select few labs.
The real breakthrough here isn't that AI got faster. It's that AI finally understood the physics well enough to be trusted with the answer.
Join the Signal
If AI tools like DONUT keep pushing scientific analysis from months down to seconds, how much do you think that changes the pace of real-world breakthroughs in batteries, catalysts, and electronics over the next few years? Share your take in the comments.
#ArtificialIntelligence #MaterialsScience #XRayScience #MachineLearning #Argonne #ScientificComputing #Nanotechnology #ScienceNews #KMManoharInsights
Explore More
Scientist watches chemistry unfold atom by atomAnother story pushing the boundaries of what X-ray science can reveal, from a very different angle.
Robots Running Laboratories in Japan: The Rise of Self-Driving ScienceA closely related shift — experiments that can plan and adapt their own next steps, with far less human hand-holding.
Laser Phase Plate: The Microscope Breakthrough Revealing Nature's Hidden ProteinsAnother imaging breakthrough built around seeing structure that used to stay frustratingly out of reach.
AI Semiconductor Boom 2026: The Chip Race Reshaping Global TechThe wider AI-and-hardware wave this kind of physics-aware machine learning is quietly riding alongside.
Randomly Mixed Atoms Arranged in Rows and Columns for Sustainable CatalysisA materials science story that leaned on exactly the kind of atomic-scale imaging tools DONUT is now making faster.
About the Author
K M Manohar is an independent Sci-Tech writer and research curator who explains breakthrough discoveries that are shaping the future of technology.
Comments
Post a Comment