Machine Learning/Nuclear Engineering and Technology 2026
Thirty-three hours of computing,
done in thirty-five seconds
- Reduced-order model (ROM)
- Proper orthogonal decomposition (POD), autoencoder, dynamic mode decomposition (DMD)
- LSTM time-series prediction
- LES (WALE) reference simulation
Working out exactly how water moves inside a reactor takes more than a day of computing. You cannot re-run that every time you nudge the design. So instead of computing the whole flow, we taught a model to carry only what matters.
- Bum Jin Cho
- Chae Hyeon Song
Min Seop Song- et al.
01 — What device is this about
Where a reactor's heat
turns into steam
A nuclear plant, in the end, boils water. The steam spins a turbine and that makes electricity. The device that carries heat from the reactor's water into a second stream of water is the steam generator.
A Small Modular Reactor is, as the name says, small. Fitting a steam generator inside a compact vessel leaves very little room, so the tubes are wound like a spring. The same space then holds a much longer tube and a much larger surface for heat to cross. That is a helical coil steam generator.
But once the tubes are coiled, the water no longer runs alongside them. It runs across them. That sideways passage is called cross flow.
02 — What goes wrong
The water swirls behind each tube
Water striking a tube leaves a swirl behind it, the way wind goes ragged past a lamp post. The trouble is that the swirl will not stay put.
The tubes shake
A vortex swinging from side to side pushes and pulls on the tube. Repeated long enough, that wears the tube out.
Heat moves unevenly
Water sweeps briskly past some spots and barely reaches others, so heat does not cross evenly either.
There is no steady beat
A swing as regular as a pendulum would at least be easy to anticipate. This one lingers on one side, then flips almost at once.
03 — Why it is hard
Getting it right
takes far too long
There is a way to compute the motion of water by chopping it into very fine pieces and following each one. It misses nothing, not even the small swirls, and the answer can be trusted.
The cost is time. Computing 0.2 seconds of this flow took 33 hours and 24 minutes. You cannot wait a day and a half every time you want to try a different tube spacing.
And the result is enormous. One picture of the flow is 22,878 numbers. There are 3,000 pictures, more than five gigabytes in all.
04 — The change of mind
We stopped handling the flow as a whole
Think of a piece of music. Writing down every wobble of the sound forty thousand times a second makes an enormous file. Yet the sound is really a handful of notes layered together. Note down which notes, and how much of each, and it becomes far smaller.
Flow works the same way. Those 22,878 numbers are not each doing their own thing; a few representative shapes are layered on top of one another. One large swirl, a smaller one, a gentler ripple, and so on.
Keep the first 50 of them and 96 % of the original flow comes back. One picture is then not 22,878 numbers but 50 numbers.
The original flow
22,878 numbers
Rebuilt from shapes
6 numbers
- Energy recovered
- 62%
- Shrunk by
- 3,813×
The first four shapes
The shapes never move. The only thing that changes with time is the single number in front of each one (the bars below).
05 — Carrying it forward
Only 50 numbers have to be guessed
From here the problem is much smaller. Nothing has to guess the whole picture, only what those 50 numbers will be a moment later.
An LSTM network does that job. It works the way a phone keyboard suggests your next word: it remembers what came before and guesses what follows. Here it guesses the next values of 50 numbers rather than the next word.
We showed it the first 0.4 seconds, then asked it for the 0.2 seconds that followed. Since the true answer was known, we could count how much it got right.
06 — Three of them, side by side
There is more than one way to pick the shapes
Part of what this study did was put three methods on the same data under the same conditions. Usually each is tried on its own, which tells you nothing about which is better.
POD
Takes the largest first. The shape holding the most energy is number one, the next is number two, and so on.
DMD
Sorts by how fast each shape oscillates, separating the slow surges from the quick tremors.
Autoencoder
Lets a neural network work out the compression for itself. No one hands it the rule.
How much they got right
| Method | Learned window | Predicted window |
|---|---|---|
| PODbest | 0.997 | 0.970 |
| Autoencoder | 0.989 | 0.948 |
| DMD | 0.997 | 0.941 |
R² measures how much of the answer was captured. At 1 the prediction is exact; at 0 it is no better than quoting the average.
POD came out ahead, because it holds on to the large swirls. The autoencoder kept the finer ripples better, and DMD drifted further off as time went on. In this case, with a narrow gap between tubes and the flow strongly confined, getting the large structures right mattered most.
07 — How much faster
33 h 24 min became 35.5 s
Once trained, the model predicted the same window in 35.5 seconds, about 3,388 times faster. That is fast enough to run alongside a plant while it operates. It is the point at which a digital twin, a computational counterpart running in step with the real plant, becomes possible.
08 — What it still cannot do
This does not replace the simulation
It drifts as it goes
The further ahead it is asked to look, the more error accumulates. Useful over a short horizon; unreliable over a long one.
Fine ripples are lost
Large swirls survive compression; the fast, small tremors are smeared out. Above 150 oscillations a second, more than half of it is wrong.
It learned one condition only
Inlet speed and tube arrangement were held fixed during training. Change them and it has to learn again from scratch.
Direction is not there yet
For now it predicts how fast the water is moving, not which way. That is the next task.
It is a flat slice
A single cut-out plane rather than the full three dimensions. Going to 3D multiplies the data many times over.
What this work achieves is not doing away with the simulation but drawing the line: where a surrogate can stand in, and where it cannot.
09 — In short
If it comes down to one line
A complex flow in 50 numbers
One picture went from 22,878 numbers to 50, and 96 % of the original still survives.
3,388 times faster
33 h 24 min of computing became 35.5 s, at R² = 0.97.
Three methods compared head to head
Same data, same conditions, same measurements, so the next person has grounds for choosing.
The limits are written down too
How far it can be trusted, and where that stops.
Source
Bumjin Cho, Chaehyeon Song, Minseo Lee, Joongoo Jeon, Minseop Song, “Efficient prediction of turbulent cross flow in helical coil steam generators of SMR via deep learning-driven reduced-order models,” Nuclear Engineering and Technology 58 (2026) 104328.
doi.org/10.1016/j.net.2026.104328The paper is openly available (CC BY-NC-ND 4.0).
Earlier presentation
NURETH-21, 2025
This research was supported by the National Research Council of Science & Technology (NST) grant by the Korea government (MSIT) (No. GTL24031-000) and the research fund of Hanyang University (HY-202200000003004).