Reactor Monitoring: Powerful AI Catches Hidden Blockages

Nuclear reactor monitoring today mostly means watching what goes into a component and what comes out of it. Researchers at Argonne National Laboratory have published a reactor monitoring design that watches what happens inside one, and the component they chose is the part of a molten-salt reactor most likely to fail.

The work pairs distributed fibre-optic temperature sensing with explainable machine learning to spot coolant channels beginning to freeze shut — while they are still only partly blocked, and while the aggregate flow, temperature and pressure readings at the inlet and outlet still look entirely normal.

The paper, “Explainable machine learning for incipient anomaly detection in compact molten salt heat exchanger with overlapping feature distributions,” appeared in Scientific Reports on 6 March 2026. The authors are Konstantinos Prantikos, Taeseung Lee, Thanh Q. Hua, Lefteri H. Tsoukalas and Alexander Heifetz, and the underlying work is supported by the US Department of Energy.

What follows is what the reactor monitoring system does, the numbers behind the headline, why the honest result is more interesting than a clean accuracy figure, and what it means for anyone building nuclear reactor monitoring for advanced designs.

Table of contents

The Reactor Monitoring Problem: Salt That Freezes at 500 Degrees
Reactor Monitoring Hardware: Sensors That Avoid the Pressure Boundary
The Reactor Monitoring Data: What “Incipient” Actually Means
The Reactor Monitoring Benchmark: Eight Models, One Winner
Explainable Reactor Monitoring Is the Novel Part
What the Reactor Monitoring System Reads: Nine Features, Not One
Reactor Monitoring Caveats a Plant Operator Should Hold Onto
Reactor Monitoring Lessons Beyond Molten Salt
From Paper to Plant: What Would Have to Happen Next
Frequently Asked Questions About This Nuclear Reactor Monitoring Research
Nuclear Reactor Monitoring: The Verdict
References

The Reactor Monitoring Problem: Salt That Freezes at 500 Degrees

Molten-salt-cooled reactors are one of the leading advanced designs, and both their appeal and their reactor monitoring difficulty come from the same property.

Why molten salt is attractive

These reactors run liquid salt at ambient pressure with a typical maximum cycle temperature around 600°C, against roughly 275°C for a conventional light water reactor. Higher temperature means more efficient power conversion, and ambient pressure means thinner piping and vessels instead of the 15 MPa design pressure a pressurised water reactor requires.

Why the same property is dangerous

Those salts have melting points typically above 450°C. Argonne’s press material puts the practical freezing risk at around 500°C, or 932°F. A coolant that solidifies at a temperature most industrial processes never reach is a coolant that can plug a pipe if a section loses insulation or accumulates impurities.

Where it plugs

The heat exchanger. A compact unit contains between 2,000 and 4,000 small parallel channels, and larger units contain more. As Argonne principal electrical engineer Alexander Heifetz frames the requirement: “If one of those channels becomes plugged, either partially or completely, we want to be able to detect it, locate it and determine the severity.”

Why conventional nuclear reactor monitoring misses it

Standard practice measures temperature, pressure and flow at the inlets and outlets. One partly blocked channel out of several thousand changes those aggregate figures by an amount indistinguishable from noise. By the time the aggregate moves, the fault is no longer incipient — and heat exchanger failure forces an unplanned shutdown that disrupts a refuelling cycle measured in years.

Property	Light water reactor	Molten-salt-cooled reactor
Maximum cycle temperature	About 275°C	About 600°C
Operating pressure	About 15 MPa	Ambient
Coolant melting point	0°C	Typically above 450°C
Channel plugging risk	Not applicable	Primary heat exchanger fault mode
Vessel wall thickness driver	Pressure	Temperature and corrosion

Reactor Monitoring Hardware: Sensors That Avoid the Pressure Boundary

The hardware half of the contribution is a heat exchanger geometry chosen because it can be instrumented for reactor monitoring, rather than an instrument retrofitted to an existing geometry.

The matrix heat exchanger

The proposed unit is a compact salt-to-salt matrix type: interleaved arrays of parallel tubes separated by divider plates. The authors are explicit about why that shape was picked — “the linearity and relative simplicity of the matrix HX geometry allows for installation of distributed temperature sensors to detect and localize faults in the individual channels.”

Sensing through the structure, not through the wall

Temperature is read at points on the divider plates rather than inside the channels. The sensors ride on mechanical support structures instead of penetrating the heat exchanger wall, which preserves structural integrity. In a pressure-boundary component, every reactor monitoring penetration is a qualification problem and a leak path, so removing them is worth a great deal on its own.

Why fibre optics rather than thermocouples

Distributed fibre-optic sensing gives “higher spatial sampling density and fewer physical penetrations through thermal hydraulic components pressure boundaries” than a conventional thermocouple array. One fibre replaces hundreds of discrete sensors and hundreds of feedthroughs.

The restraint that makes it work

The modelled design samples 1,428 gauge points at the front and back of the tubing manifold — far fewer than the fibre could physically provide. The authors explain why: reading every possible point “would result in high rate of false alarms,” because partial plugging shifts local divider-plate temperature by only a few degrees, barely above the sensor’s own uncertainty. Reactor monitoring sampling density was deliberately traded against false-alarm rate.

The Reactor Monitoring Data: What "Incipient" Actually Means

This is a modelling study, and being precise about that is essential to reading the reactor monitoring result correctly.

Synthetic, not experimental

Thermal and pressure fields were generated with COMSOL heat transfer and pipe modules across 36 steady-state cases, with realistic sensor noise added afterwards. Plugging was simulated by reducing flow in selected channels. No salt was frozen in a laboratory for this paper.

The three severity classes

Faults are graded rather than binary, which is what lets the reactor monitoring system report severity instead of just raising an alarm. Class 1 is the mildest, a 20% reduction in channel flow; class 2 is 40%; class 3 is 60%. Class 0 is a fully open channel.

The imbalance

The full dataset contains 25,704 instances, of which 24,772 are open channels and 932 are plugged — a minority-to-majority ratio of 3.76%, split as 1.41% class 1, 1.20% class 2 and 1.02% class 3. The held-out test set holds 4,998 instances, 4,749 open and 249 plugged across the three severities.

The overlapping distributions in the title

Because heat diffuses through the divider plates, a mildly plugged channel and a healthy one produce temperature readings that genuinely overlap. The dataset is non-separable by construction. That is the whole point of the paper — it is a study of the regime where detection is hardest, not the regime where it is easy.

Test set composition (4,998 instances)

Class 0, open channels 4,749

Class 1, 20% flow reduction 82

Class 2, 40% flow reduction 84

Class 3, 60% flow reduction 83

The Reactor Monitoring Benchmark: Eight Models, One Winner

Eight supervised classifiers were compared for the reactor monitoring task: logistic regression, k-nearest neighbours, Gaussian naive Bayes, a support vector machine, decision trees, random forest, a feed-forward neural network and XGBoost.

Why accuracy was not the metric

With 95% of instances in one class, a model that predicts “open” for everything scores about 95% accuracy and is worthless. The authors say so directly, and report confusion matrices, F1-scores, precision-recall curves and area under the precision-recall curve instead. For context, the baseline AUC-PR for each minority class is 0.02.

The error that matters most

The critical failure is calling a plugged channel healthy. XGBoost produced the fewest such errors, 36, against 39 for decision trees, 58 for random forest and 62 for the support vector machine. That is the reactor monitoring number a plant operator would be judged on.

Per-class results

XGBoost led on F1 for class 3 at 0.96 and was competitive on class 2 at 0.82, where decision trees edged it at 0.85. The honest column is class 1, where the best F1 across all eight models is XGBoost’s 0.51 — a coin flip.

Model	F1 class 1 (20%)	F1 class 2 (40%)	F1 class 3 (60%)	Missed plugs
XGBoost	0.51	0.82	0.96	36
Decision tree	0.49	0.85	0.94	39
Random forest	0.48	0.82	0.93	58
Support vector machine	0.37	0.80	0.94	62
Feed-forward network	0.31	0.48	0.80	54
k-nearest neighbours	0.16	0.59	0.91	102
Logistic regression	0.00	0.41	0.90	134
Gaussian naive Bayes	0.04	0.13	0.45	150

XGBoost F1-score by plugging severity

Class 1, 20% flow reduction 0.51

Class 2, 40% flow reduction 0.82

Class 3, 60% flow reduction 0.96

Class 0, open channel 0.99

Reading that chart correctly

Detection quality scales with severity, which is the expected and physically sensible result. A nuclear reactor monitoring system built on this would reliably catch a channel that is 60% blocked, usually catch one at 40%, and be unreliable at 20%. That is still earlier than inlet-outlet instrumentation, which catches none of them.

Explainable Reactor Monitoring Is the Novel Part

Plenty of papers apply gradient boosting to sensor data. The reactor monitoring contribution here is what sits on top of the classifier.

Shapley values, and their weakness

SHAP values quantify each feature’s contribution to a prediction. The authors identify the primary outlet distributed temperature as the dominant predictor, “exerting a significantly greater impact than all other features combined” — but they note that Shapley values impose a total ranking even when differences are statistically meaningless, which “can lead to overconfident interpretations of model behavior.”

Adding partially ordered sets

To handle that, the team constructs partially ordered sets from the SHAP values and renders them as Hasse diagrams. A POSET can say that two features are incomparable rather than forcing one above the other, so the explanation distinguishes confidently ranked features from ambiguous clusters.

What the diagrams reveal

The structure changes with severity. Under normal operation, gauge proximity and divider plate number sit clearly below the dominant temperature feature. At 20% plugging the secondary level collapses into one large ambiguous cluster, meaning “early-stage plugging events depend on subtle signals captured collectively rather than individually.” At 60% plugging, gauge proximity re-emerges as a distinct second-ranked feature.

Why a regulator will care

A safety case needs to state what the reactor monitoring model relies on and how confident that statement is. An explanation that admits ambiguity is more defensible in front of a regulator than one that asserts a false precision, and this is the first part of the work that looks designed for licensing rather than for publication.

What the Reactor Monitoring System Reads: Nine Features, Not One

An easy misreading of the headline is that reactor monitoring here means watching temperature. It watches nine variables, and four of them are not measurements at all.

The contextual features

Gauge location along the fibre, gauge proximity to a primary or secondary channel, measurement number and divider plate number describe where a reading came from rather than what it says. Divider plate number and gauge location together act as the coordinates that locate a fault, which is how the system answers Heifetz’s “locate it” requirement.

The physical measurements

Primary inlet and outlet distributed temperatures from the fibre, plus conventional thermocouple readings at the combined primary and secondary outlets, plus a secondary-side pressure drop from a pressure transmitter. The new distributed measurements complement the conventional instrumentation rather than replacing it.

The features they threw away

Fourteen predictors were evaluated and nine retained after ablation. Dropped were the thermocouple inlet temperatures on both sides, the primary-side pressure drop and both loop flow rates — in other words, most of the signals a conventional nuclear reactor monitoring system is built on.

Why no single feature is enough

Relying on the dominant temperature feature alone was “insufficient to achieve adequate predictive performance,” and a three-feature subset performed poorly. The dominant feature detects; the contextual features localise and grade. That coupling is the practical design lesson.

Reactor Monitoring Caveats a Plant Operator Should Hold Onto

Four limits sit between this paper and a reactor monitoring system anyone would licence.

It is synthetic data

COMSOL fields with added noise are not molten salt. Real installations bring corrosion, impurity accumulation, insulation degradation, fibre drift under radiation and thermal ageing — none of which appear in a steady-state simulation. Moving a model from simulated to measured data is a transfer learning problem in the strict sense, and the usual outcome is that performance falls before it recovers.

Steady state only

All 36 cases are steady state. Real plants start up, shut down, ramp and cycle, and the transients are exactly when local temperature gradients are largest and a plugging signature is most likely to be masked.

Class 1 is not solved

An F1 of 0.51 on the earliest fault stage means roughly half of the mildest plugs are missed or misgraded. The paper is honest about this; a vendor datasheet derived from it might not be.

Sensor qualification

Fibre in a 600°C salt environment inside a nuclear pressure boundary is its own multi-year qualification programme, independent of any machine learning. In reactor monitoring the instrumentation, not the model, is the long pole.

Reactor Monitoring Lessons Beyond Molten Salt

The transferable reactor monitoring ideas here are not nuclear-specific, and they are worth extracting.

Instrument the structure, not the flow

Putting sensors on supporting structures rather than through pressure boundaries is a pattern that applies to any high-integrity system where penetrations are expensive to qualify. You give up signal fidelity and buy back an entire category of failure.

Grade the fault, do not just flag it

Reporting 20%, 40% or 60% rather than “fault” turns an alarm into a maintenance decision. A binary detector tells an operator to investigate; graded reactor monitoring tells them whether to schedule the work for the next outage or the next hour.

Design the sampling rate for the false-alarm budget

Deliberately reading 1,428 points instead of every point the fibre offers is the most quietly useful decision in the paper. More sensing is not better reactor monitoring when the incremental channels sit below the noise floor, and this is a lesson most industrial predictive analytics programmes learn the expensive way.

Explanations that admit uncertainty

The SHAP-plus-POSET approach is portable from reactor monitoring to any regulated domain where a model’s reasoning has to be stated. As Argonne’s Prantikos puts it, the appeal of the approach is that “our approach can detect barely distinguishable patterns in large volumes of data that would be challenging for a human to inspect” — and, he adds, “AI is always on guard. It doesn’t sleep.” Teams standing up equivalent capability internally usually find the bottleneck is ML model development discipline rather than algorithm choice.

From Paper to Plant: What Would Have to Happen Next

Nothing in this result is close to a deployed product, and the path between the two is long enough to be worth setting out explicitly.

A salt loop with real fibre in it

The obvious next experiment is a bench-scale molten salt loop with a matrix heat exchanger, real distributed fibre on the divider plates, and deliberately induced plugging. That produces measured rather than simulated distributions and tells you immediately how much of the classifier’s performance survives the move.

Radiation and thermal ageing data

Optical fibre in a reactor environment darkens under radiation and drifts with thermal cycling. A monitoring model that assumes a stable sensor response will degrade quietly as the fibre ages, which is worse than failing loudly. Characterising that drift is a separate multi-year programme.

Transient operating envelopes

Every case in this study is steady state. Adding startup, shutdown and load-following data roughly multiplies the problem: the model has to distinguish a plugging signature from the far larger temperature gradients a ramp produces, and those gradients are where the signal is weakest.

A safety case, not just a model

Finally, a regulator will ask what happens when the classifier is wrong in each direction, what the operator is instructed to do, and how the model is revalidated after any change to the plant. That is where the explainability work pays off, and it is also where most industrial machine learning projects discover that the model was the easy part.

Stage	What it proves	Status
Simulated fields, added noise	The signal exists in principle	Published, March 2026
Bench salt loop with real fibre	The signal survives real physics	Not reported
Irradiated and aged sensors	The signal survives service life	Not reported
Transient operating data	The signal survives ramps	Not reported
Regulatory safety case	Operators may act on it	Not started

Frequently Asked Questions About This Nuclear Reactor Monitoring Research

Has this been tested in a real reactor?

No. The study uses high-fidelity synthetic data generated with COMSOL simulations, with sensor noise added to approximate real measurements. It is a reactor monitoring design and feasibility study, not an operational trial.

Which model performed best?

XGBoost. It produced the fewest missed plugs at 36 and the best F1-scores on the most severe fault class at 0.96, with tree-based methods generally outperforming the alternatives.

How early can it detect a blockage?

Reliably at 60% flow reduction, usually at 40%, and unreliably at 20%, where the best F1 across all eight models was 0.51. All three are earlier than inlet-outlet reactor monitoring, which does not register any of them.

Why does it use fibre optics?

Distributed fibre-optic sensing gives far higher spatial sampling density than thermocouples and needs far fewer penetrations through the pressure boundary, which matters more in a nuclear component than the raw measurement quality does.

Who funded and published it?

The work comes from Argonne National Laboratory with Purdue collaboration, supported by the US Department of Energy, with the Advanced Research Projects Agency-Energy cited in Argonne’s own release. It was published in Scientific Reports on 6 March 2026.

Nuclear Reactor Monitoring: The Verdict

This is a careful piece of reactor monitoring research that resists the temptation to report a flattering number. The design choices — instrument the divider plates, sample fewer points than the fibre allows, grade severity rather than flag faults, and explain predictions in a way that admits when the explanation is uncertain — are all aimed at a system a regulator could eventually accept rather than a demonstration that looks good in a chart.

The honest caveat is that the strongest claim available today is that a simulated heat exchanger with simulated faults can be diagnosed by a gradient-boosted classifier at severities above roughly 40%. That is a real advance over measuring only what enters and leaves the component, and it is a long way from a licensed nuclear reactor monitoring product. The next paper to look for is the one with salt in it. Readers following the wider energy story may also want our coverage of California’s new rules on AI data centre energy and water use, which is where much of the demand for advanced reactors is coming from.