The paperwork
A plant has to tell the regulator how much nuclear fuel it holds and where it went. Simple with a few labelled fuel rods. Hard with hundreds of thousands of identical balls that never stop moving.
Fuel handling · identity
A pebble bed reactor runs on graphite balls that circle through the core for years. They are made to be identical. Everyone else either cuts a code into the shell and hopes it survives, or puts an X-ray scanner in the fuel line. This station reads the graphite surface itself: nothing added, nothing cut, names the ball, counts its passes and spots a cracked shell from the same glimpse. The method is built and benchmarked to the physics of the real thing.
Every one of them is a real render of a different ball. Go on, guess.
The loop
Balls enter the top, trickle down making heat, leave through a chute one at a time, get measured, and go back round for another pass. About six passes over three and a half years, then they retire. One ball reaches the station roughly every ninety seconds.
The object itself
Three views of the same ball, drawn to the published dimensions. The third is the one the strongest work in this field fingerprints, and it is not the surface.
Renders to published dimensions: 60 mm sphere, 5 mm fuel-free shell, about 18,000 particles on a 910 micrometre minimum spacing. Sources: Xe-100 core design report, ANL/SS-25/19, Fang and Di Fulvio 2023.
Value
A plant has to tell the regulator how much nuclear fuel it holds and where it went. Simple with a few labelled fuel rods. Hard with hundreds of thousands of identical balls that never stop moving.
Knowing each ball's history means sending it round for exactly the right number of passes instead of guessing from averages. More energy out of the same fuel. How much more is a question for the reactor designer, not for us.
Graphite goes brittle under heat and radiation. A ball cracks, sheds a chunk, and the chunk blocks the chute where balls pass single file. Argonne put it plainly in their November 2025 report: the larger concern is not the radiation from a cracked ball, it is fragments impeding the flow of pebbles through the reactor system.
A four unit Xe-100 is 320 megawatts of electricity. Every day the chute is blocked is about 7,700 megawatt hours that never get sold. Price that at your own tariff. Spotting the crack before it breaks is worth more than the paperwork, and the same glimpse that names the ball also sees the crack.
The trick
All the balls are the same part, so most of what a camera sees is shared. Take away the family likeness and what remains belongs to one ball only. The third panel is exactly what the second one added to the first, in the same colours, with the contrast raised six times because at its real strength you would see almost nothing. That faintness is the difficulty, and it is why a person cannot do this job.
Every new ball scores near zero against every earlier one. That gap is the product.
The geometry
A sphere arriving in a random orientation gives a camera nothing to aim at. Two photos of the same ball usually show different parts of it, and share about a twentieth of the surface. Not enough to compare. This is why the incumbent's approach, which needs a known region on a known kind of object, has nothing to grip.
Press "Take one photo" twice to see two separate photos of the same ball.
Roll each new ball once on a turntable and photograph it all the way round. That happens at fuel receipt, where a fixture is easy and only a couple of hundred balls arrive a day. Afterwards one glimpse in the hot chute is enough, because wherever the ball lands, that patch is already on file.
The station
One screen beside the discharge chute. It recognises a ball it enrolled months ago, counts its passes, flags a cracked shell from the same glimpse, and asks for another look when it is not sure. Asking again is a feature: a wrong name in a nuclear record is worse than no name.
Ready
Press "Read next ball".
Gallery of 240 enrolled balls, each with its own pass history. Scores are drawn from the measured distributions below. In 1,740 benchmark comparisons, "wrong name" never fired once.
What we measured
Thirty balls, each enrolled by rolling it all the way round, then read again from one random angle. Every read was scored against every template. This is the whole dataset, not a selection. It is a surface method, so it is measuring the thing the X-ray work never has to worry about.
Right ball named first time, 60 reads out of 60. Sixty reads puts the true rate between 94 and 100 percent.
Dust film, a quarter of the surface abraded, blur, and light twenty degrees off axis: 59 of 60. The one miss was a retake. No ball was ever given the wrong name.
The worst right-ball score was 0.391. The best wrong-ball score was 0.117. Nothing landed in between, which is why the station can ask for another look rather than guess.
Benchmarked on spheres simulated to published graphite roughness, wear per pass and camera noise.
The competition
Their patents require a stored template for the object's class that defines where to look: a region on a watch bezel, a corner of a label, the face of a chip. One item at a time, checked against its own record, usually with a phone.
A grey sphere arriving in a random orientation gives that nothing to aim at. There is no bezel and no corner.
We start from the population. Because every ball is the same part, the family likeness can be measured and removed, which is the step a one-at-a-time checker has no reason to perform.
And because the inventory is closed and finite, a name has to be consistent with every other ball's name, not just plausible on its own.
Their 99.85 percent was an objective written into a funding application, not a measured result.
Kairos built a line that inspects every pebble, then laser-engraves three redundant codes half a millimetre into the graphite shell, and hopes the codes survive years of tumbling. That is a cut into the fuel-free shell on every ball, a reader that trusts whatever the code says, and nothing at all for a ball whose code has worn or chipped. Residue reads the ball itself. No mark, no cut, no extra process step, and it works on every pebble ever made, including theirs. The same glimpse that names the ball sees the crack a damaged code would hide.
The national labs are testing coils that tell solid from annular pebbles and survive the heat and dust. They propose cracks and voids as signatures for identity. A coil image is just another picture of the surface. The engine on this page runs on it unchanged, which is why the station is built to take a frame from a camera or a coil.
Alitheon won DOE SBIR Phase I award 201910, 195,828 dollars, topic C56-40n, in 2023, titled "FeaturePrint: Computer Vision for Pebble Bed Reactors' Material, Control, and Accounting". The abstract lists reaching 99.85 percent accuracy among its objectives. There is no public Phase II for that topic and no independent test of their method, and their software is not available.
Ming Fang and Angela Di Fulvio at Illinois identify pebbles by the three dimensional pattern of the fuel particles sealed inside them, read with X-ray CT. Spherical harmonics narrow a library of 100,000 down to 100 candidates, then a rotation-only alignment picks the one. They report 100 percent identification across their test sets, under arbitrary rotation and heavy noise, in under 50 seconds a pebble including imaging. Their library and test pebbles were simulated and their physical scan used a mock-up compact rather than a pebble, so both approaches stand on simulated pebbles today.
The particle pattern is sealed under 5 mm of graphite. Tumbling, abrasion and dust cannot touch it, which is exactly the risk hanging over any surface method, ours included.
An industrial CT scanner at 200 kilovolts in the discharge path, plus the licensing and shielding that comes with it. They also flag that spent pebbles may saturate the detector, which is the hot path this has to work in.
It is no longer whether a pebble can be identified. It is whether it can be identified with a camera and a coil instead of a CT scanner, at a hundredth of the cost, on every ball rather than a sample.
Fang and Di Fulvio, Rotation-invariant rapid TRISO-fueled pebble identification, Annals of Nuclear Energy, 2024. See also Stringer, van der Ende and Anghel, identification from pairs of X-ray radiographs, Nuclear Instruments and Methods A, 2024.
The bar
A reactor reads a ball about every ninety seconds, which is 372,300 reads a year. At 99.85 percent that is roughly 558 wrong names a year, about one and a half a day. Below is a year. A red dot is a day on which at least one ball went into the record under the wrong name.
Status
The map is drawn, the method is built, and it holds under every condition the physics allows.
Same engine
The station takes a frame and returns a name. A coil image works the same as a photo, which matters because the labs are already building the coil.
A camera on the chute, or frames pulled from a feed that already exists. Several frames of one ball beat one, because each adds surface.
A coil image instead of a photo. Argonne's November 2025 work reads cracks and voids through the outer graphite, works at 400 degrees, and is unaffected by the ball's own radiation.
The ball's own heat as it leaves the core. Fast to capture, shallow, useful as a second opinion rather than a primary.