Machine learning is opening new possibilities for monitoring Small Modular Reactors using neutron noise

The transition away from fossil fuels is one of the main challenges of this century. Electricity grids that have long depended on coal and natural gas need to be significantly decarbonised within a timeframe measured in decades, not generations. Renewable sources such as wind, solar, and hydropower are expanding rapidly, but their variable and location-dependent output means that alternative low-carbon generation will remain essential for grid stability at various locations. Nuclear power is the most established technology capable of filling that role on a large scale, and it is getting attention across many national energy policies as a result. Beyond electricity generation, nuclear reactors – and Small Modular Reactors (SMRs) in particular – also offer high potential for process heat, district heating, and hydrogen production, opening pathways to decarbonise heavy industries such as steel and petrochemicals.

Within this renewed interest, a particular category of reactor design has emerged as a very attractive option: SMRs. SMRs are defined as reactors with an electrical output of up to 300 MWe that are designed in smaller modules to be fabricated in factories and transported to the deployment site.¹ Compared with large conventional reactors, which typically generate 1000 MWe or more and require decade-long construction programmes, SMRs offer a different value proposition. SMRs can be deployed in regions where the electricity demand does not justify a full-scale plant, or where grid infrastructure limits the size of power generation that can be integrated. The smaller power output of SMRs allows for improved passive safety systems based on natural circulation and gravity rather than active pumping, i.e., without external power or operator intervention. Their compact design results in reduced construction times and reduced capital investments. However, whether SMRs will achieve economic competitiveness with large reactors at scale remains an open question, and one that will depend considerably on the ability of manufacturers to industrialise production and reduce per-unit costs.²

Reading neutrons distribution as a diagnostic tool

Nuclear reactors are never perfectly static systems. During normal operation, coolant flows through fuel channels, mechanical structures undergo small-amplitude vibrations, and the reactor vessel experiences minor oscillations. All of this physical activity induces fluctuations that leave a measurable trace: stationary, small-amplitude time-fluctuations in the neutron flux distribution around its mean value, known as ‘neutron noise’, even when the reactor is at steady-state conditions.

When an anomaly develops, such as an abnormally vibrating fuel assembly, a change in local coolant density, or a degrading mechanical component, the neutron noise pattern changes in characteristic ways.³ As these changes appear as soon as a physical disturbance begins to develop, neutron noise has the potential to detect anomalies at an early stage, before they escalate into more serious operational or safety-related issues. This gives operators the opportunity to intervene early and reduce the risk of unwanted events.

The central question driving this research is whether those changes can be detected, interpreted, and localised automatically, without interrupting reactor operation and without relying on dense arrays of instrumentation that practical reactors cannot accommodate. As part of the ANItA research centre – a Swedish national competence centre for nuclear power technology – the possibility to use neutron noise for anomaly detection in water-cooled SMRs was investigated.

Neutron noise diagnostics has a long history, dating to oscillator experiments carried out at Oak Ridge National Laboratory in the late 1940s.⁴ The underlying principle is physically well-motivated. Because neutrons propagate by fission and scattering reactions throughout the entire core volume, a localised perturbation anywhere in the core induces a measurable response at all detector locations, including those far from the noise source. This non-local sensitivity is what distinguishes neutron-based monitoring from, for example, temperature measurements, which are inherently local and do not necessarily carry information about conditions at distant points in the system.

The spatial distribution of the neutron noise carries information that is critical for localisation. The neutron noise consists of two components: the ‘point-kinetic component’, which has the same spatial shape as the static neutron flux distribution and therefore carries no spatial information about the perturbation source, and the ‘deviation from point-kinetics component’, which carries characteristic spatial information tied to the perturbation’s location and nature. It is the latter component that diagnostics must extract and interpret, if one wants to retrieve the position of a local perturbation.⁵

For decades, neutron noise diagnostics was developed and validated primarily for large commercial light-water reactors, with cores on the order of several metres in both height and diameter.⁶ In these large systems, detectors placed at different positions within the core record measurably different responses to the same perturbation. This spatial diversity is precisely what makes it possible to trace a disturbance back to its origin.

The smaller the core, the weaker the interpretable signal

The growing interest in SMRs introduces a significant complication. The reduced core dimensions of SMRs fundamentally alter the neutron noise characteristics. In a smaller core, the neutron noise distribution becomes more spatially similar to static flux in response to localised perturbations. The result is that the point-kinetic component becomes increasingly dominant, while the deviation from point-kinetics is suppressed.⁷ In the limiting case of a very small core, all detector signals become nearly identical to static flux regardless of where the perturbation is located, making traditional spatial localisation methods ineffective. The practical consequence of this phenomenon is visible in Fig. 1. In a large reactor, the discrepancy between neutron noise recorded at different positions and its respective point-kinetic component varies substantially across the core, which presents a rich source of diagnostic information. In a small reactor, that variation is significantly suppressed, and detectors across the core record signals that are nearly identical to the static behaviour.⁸

This fundamental physical constraint is reinforced by practical instrumentation limitations. SMRs are designed for simplicity and economy of operation, which typically means fewer in-core detector positions than large reactors. With fewer measurement points already receiving a spatially near point-kinetic signal, extracting reliable localisation information becomes substantially more difficult. These two factors together mean that diagnostic methods developed for large reactors cannot be directly transferred to SMRs. New approaches, adapted to the specific physical and instrumentation constraints of compact cores, are required.

Machine learning as a diagnostic tool

The central insight of this work is that, even in an SMR core, the spatial characteristic signal of a perturbation is significantly weakened but not absent. It remains encoded in the subtle differences between the neutron noise distribution recorded across the core, even when those differences are small relative to the dominant point-kinetic background.

Machine learning has become one of the most remarkable developments in modern computing. It underlies tools that recognise faces, translate languages, identify tumours in medical scans, and forecast weather. In each application, the core principle is the same, i.e., instead of programming a system with explicit formulas, it is trained on large numbers of examples until it learns to identify the patterns that matter. After learning the pattern, the trained model can apply that knowledge to new, previously unseen cases.

The same principle is applicable to reactor diagnostics.⁹ Identifying where an anomaly is located inside a reactor core, given only a set of neutron detector readings, can be considered as fundamentally a pattern recognition problem: what does the neutron noise look like when a perturbation is present at a particular location, and can a model learn to recognise those signatures reliably?

The main practical obstacle is data. A machine learning model requires large numbers of training examples to learn effectively. Real reactor anomalies are fortunately rare, which makes them difficult to observe under controlled conditions. The solution adopted here is to generate training data computationally, using detailed physics-based simulations of the SMR core. Each simulation produces the neutron noise distribution corresponding to a specific anomaly at a specific location within the core, building a large library of labelled examples. The model is trained on this library and subsequently tested on new simulations it has not previously encountered.

Toward continuous and automated core monitoring

The ongoing results of this work show that the studied approach works when applying machine learning models. Despite the physical constraints of compact reactor cores and the limited number of detectors available in realistic SMR designs, the use of complex machine learning architectures to train models successfully identifies the location of anomalies within a detailed three-dimensional model of the SMR core. Importantly, the models also perform reliably on scenarios they were not explicitly trained on, which is an essential property for any diagnostic tool intended for practical use.

Perhaps the most practically attractive feature of this approach is that it requires no modifications to the reactor. Neutron noise is always present during normal operation and is already recorded by the detectors installed in the reactor. The diagnostic system simply makes use of signals that exist by analysing them continuously and automatically, while the reactor continues to operate as normal. There is no need for additional in-core hardware, scheduled interruptions, or any form of intervention in the core. However, minor changes of the data acquisition system may be required to achieve the necessary sampling frequency, analogue-to-digital conversion, and signal filtering.

As SMRs move toward commercial deployment over the coming decades, the ability to monitor their cores reliably and continuously will be an important part of ensuring their safe operation. This work is a step in that direction, demonstrating that even in the more challenging conditions presented by a compact reactor core, the information needed to detect and localise anomalies is accessible, and that machine learning provides an effective means of extracting it. The findings also offer guidance for instrumentation design decisions, such as the optimal number and placement of neutron detectors. This is especially useful at the stage where SMR core are still in the design phase, providing an opportunity to incorporate these insights before designs are finalised.

References

1. IAEA, Advances in Small Modular Reactor Technology Developments A Supplement to: IAEA Advanced Reactors Information System (ARIS). 2022, Vienna, Austria: Vienna, IAEA
2. Pioro, I.L., Handbook of Generation IV Nuclear Reactors. 2016, Woodhead Publishing
3. IAEA, Advanced Surveillance, Diagnostic and Prognostic Techniques in Monitoring Structures, Systems and Components in Nuclear Power Plants. 2013: No. NP-T-3.14, Vienna, IAEA
4. Hoover, J.I., et al., Measurement of Neutron Absorption Cross Sections with a Pile Oscillator. Physical Review, 1948. 74(8): p. 864-870
5. Pázsit, I. and C. Demazière, Noise Techniques in Nuclear Systems. Handbook of Nuclear Engineering, ed. D.G. Cacuci. 2010, Boston, MA: Springer.
6. NEA. Reactor Noise. in SMORN-III. 1981. Tokyo, Japan: OECD Publishing
7. Demazière, C., A. Rouchon, and A. Zoia, Understanding the neutron noise induced by fuel assembly vibrations in linear theory. Annals of Nuclear Energy, 2022. 175(109169)
8. Hussein, S.M. and C. Demazière, Machine learning-based noise diagnostics for water-cooled SMRs: proof of principle on 2-dimensional systems. Progress in Nuclear Energy, 2025. 189: p. 105950
9. Kollias, S., et al., Machine learning for analysis of real nuclear plant data in the frequency domain. Annals of Nuclear Energy, 2022. 177(109293)

Acknowledgements

This research has been conducted within ANItA – Academic-industrial Nuclear technology Initiative to Achieve a sustainable energy future, which is financed by Swedish academia, the Swedish nuclear industry and the Swedish Energy Agency.

Please note, this article will also appear in the 26th edition of our quarterly publication.