Artificial intelligence in nuclear technology – between potential and responsibility
There are different approaches to AI applications worldwide, some of which are at very different stages of development. In order to take a differentiated view of the potential and limitations of AI in nuclear technology, it is useful to divide the areas of application into different categories (see drop-down menu below). In the following, specific examples are given to provide an overview of the current status of AI applications in practice and the corresponding activities on the part of authorities and regulators.
In the field of engineering and design, AI enables model-based development and simulation of complex nuclear systems. Here, so-called digital twins serve as virtual representations of individual components or subsystems, such as the reactor core or reactor coolant system, and allow the analysis of operating behaviour as early as in the planning phase. AI-based methods are also used in the optimisation of reactor components and the design of fuel architectures. By combining physical models with data-driven approaches, design decisions can be made on a more informed basis and technical innovations can be accelerated.
In the field of diagnostics, artificial intelligence can support the early detection of faults and material degradation in nuclear installations. By analysing extensive sensor data, such as from vibration, temperature or radiation measurements, anomalies can be identified that indicate corrosion, cracking or other structural changes. AI models recognise patterns and deviations that would be difficult to identify using conventional analysis methods and enable continuous condition monitoring of safety-relevant components such as pipes, pumps or transducers.
AI-based forecasting methods can be used to predict operating conditions and technical developments within nuclear systems. They are intended for use in modelling and predicting temperature curves, pressure developments or the wear behaviour of materials. Furthermore, they can support load forecasting in power grids, especially in combination with other energy sources.
In critical situations such as accidents or emergencies, AI is intended to contribute to real-time analysis and decision support. Rapid evaluation of available data should enable deriving recommendations for action, which in turn will improve emergency management. Here, digital twins represent the current state of a real system. They can thus be integrated into simulations or used as a simulation environment to analyse hypothetical events and response strategies. AI is also playing an increasingly important role in the field of cybersecurity, for example in detecting and defending against attacks on digital infrastructures of nuclear installations.
AI can contribute to optimising operational processes and increasing efficiency in nuclear installations. AI-based optimisation assistance systems are already being used today to support control processes such as regulating coolant flows or fuel utilisation. Learning systems can be used to improve processes and adapt them to changing conditions. This should not only increase economic efficiency, but also enhance operational safety by reducing human error and making more targeted use of technical resources.
AI applications in technology development
Various applications of artificial intelligence are being used in development projects around the world. So-called grey-box digital twins, for example, are hybrid models that combine physical and data-driven approaches to realistically simulate nuclear systems. In the development of nuclear reactors, they are intended to enable precise simulation of plant behaviour under various operating conditions. AI-supported components within these twins are used for error correction, real-time prediction and risk analysis. The integration of real-time data allows design decisions to be validated and safety-relevant aspects to be taken into account at an early stage of development.
In general, it can be said that machine learning is increasingly being used to simulate complex physical processes in reactor development. Neural networks, especially deep neural networks, can, for example, simulate thermodynamic processes, flow behaviour and material reactions with a high degree of accuracy. These simulation models are intended to enable different design variants to be compared and the effects of technical parameters on the safety and reactor performance to be assessed.
One example is the Xe-100 by the US company X-energy, a concept for a gas-cooled high-temperature reactor with modular design. In the development process of this reactor, AI is used in several areas: AI models are helping to optimise the reactor design in terms of passive safety, transportability and assembly speed. Moreover, AI is used to simulate thermal processes and analyse material behaviour, which, according to the manufacturer, will reduce the need for physical prototypes and enable more substantiated design decisions. AI is also used for quality control and process control in the manufacture of the required TRISO-X fuel, which is designed for high temperatures.
The Aurora SMR by Oklo Inc. is a compact reactor design that, according to its developers, is designed for nuclear fuel recycling and flexible site selection. AI is being used in the development process of this reactor to optimise the coupling of fuel production and reactor design. In addition, the developer is relying on AI-supported models to analyse site data and take regulatory requirements into account at an early stage. The aim is to speed up licensing processes and adapt the technical design of the reactor to different boundary conditions.
AI applications in reactor operation
During operation, AI systems can contribute to condition monitoring and fault detection. For example, condition monitoring systems can analyse sensor data to detect corrosion, material fatigue or leakages at an early stage. Studies show that neural networks are capable of classifying levels of degradation and proposing maintenance measures in a targeted manner.
One example is the AIMD system by KHNP (Korea Hydro & Nuclear Power), which uses Big Data and AI to monitor the status of over 12,000 components in 26 reactors. It combines rule-based diagnostics, frequency analysis and machine learning to detect anomalies at an early stage and recommend maintenance measures. The results are visualised in 3D models to support decision-making.
Large language models (LLMs) are also increasingly being used in nuclear technology. These are AI systems that have been trained on extensive text data and are capable of understanding and generating language. In technical applications, they can be used to structure complex information, answer questions or support decision-making, for example through interaction with operating personnel.
The Korea Atomic Energy Research Institute (KAERI) is developing Atomic GPT, an AI-based assistance system to support reactor operation. At its core is an agent system based on an LLM that assists operating personnel in a reactor simulator in performing complex tasks. The aim is to improve human-machine interaction in the control room and facilitate decision-making through context-aware suggestions. The system is being tested as part of the development of advanced reactor concepts and is intended to contribute to quality assurance in plant management in the long term.
The US company Nuclearn.ai is developing a specialised AI platform tailored to the requirements of nuclear installations. The aim is to use AI to support complex technical and regulatory processes. The platform is intended to be used, among other things, for the creation and evaluation of status reports, trend analyses, safety documentation and regulatory communication. According to the manufacturer, the solution is designed for use in high-security environments and meets regulatory standards and export control requirements.
The Russian company RASU JSC has developed an AI-based assistance system to evaluate large amounts of technical data and also contribute to decision-making in the control room. The aim is to increase operational safety through early detection of deviations. AI models are used, among other things, to analyse process data, visualise complex relationships and derive recommendations for action. In mid-2025, the system was integrated into Unit 6 of the Novovoronezh nuclear power plant in Russia.
In addition to supporting operation, AI is also playing an increasingly important role in the maintenance of nuclear installations. Among other things, the automated evaluation of test and sensor data is intended to make inspection processes more efficient and enable safety-relevant changes to be detected at an early stage. In this context, the French company Framatome has developed AI-based methods for non-destructive testing. These are used, for example, in the inspection of reactor pressure vessels and combine automated ultrasonic, eddy current and visual testing methods. The aim is to precisely identify material changes while reducing radiation exposure for personnel.
China National Nuclear Power (CNNP) also uses AI specifically to support maintenance and operational management. With the I-Nuclear platform, a digital service concept has been developed that provides AI-supported tools for commissioning, technical analysis and maintenance of nuclear installations. The platform is intended to help make safety-relevant processes more efficient, improve the quality of technical assessments and support operating personnel with digital assistance systems. Its modular structure allows different reactor types to be simulated flexibly.
A concrete example of AI-based forecasting is the prediction of moisture carryover (MCO) in boiling water reactors. MCO can lead to corrosion on turbine blades and increased radiation exposure from cobalt-60. The US software developer Blue Wave AI Labs has created an AI model that uses historical operating data to predict MCO values. This should enable targeted measures to prevent degradation and reduce personal radiation doses.
The development work carried out by Westinghouse is also worth highlighting: With the HiVE system and LLM bertha, the US company has developed a nuclear-specific AI platform designed to support the entire life cycle of reactors – from design and licensing to manufacturing, construction and operation. In collaboration with Google Cloud, a proof-of-concept was implemented in which modular construction packages for AP1000 reactors were automatically generated and optimised. The platform uses technologies such as Vertex AI, Gemini and BigQuery to accelerate design processes and improve the data-based operational management of existing plants. In addition to technical development, the system is also designed to support maintenance planning, inspection processes and digital user guidance in nuclear installations.
AI and safety
With the increasing use of AI in nuclear practice, issues of safety, licensing and regulatory assessment are also coming into sharper focus since AI systems can intervene deeply in technical processes and influence safety-relevant decisions. This not only poses new challenges for regulatory authorities but also raises questions for safety research and the assessment of technical concepts. The following outlines approaches taken by regulatory institutions and international bodies to deal with these developments.
For example, the US Nuclear Regulatory Commission (USNRC) has published a comprehensive AI strategy for the years 2023 to 2027. This strategy pursues five goals:
- Ensure readiness for regulatory decision-making
- Establish an organisational framework to review AI applications
- Strengthen and expand AI partnerships
- Cultivate an AI-proficient workforce
- Pursue use cases to build an AI foundation.
In addition, the USNRC, together with regulatory authorities from Canada and the United Kingdom, has published a paper setting out basic requirements for the safe use of AI in nuclear technology, such as transparency, robustness, ethical responsibility and consideration of human and organisational factors. These principles are intended to serve as guidance for developers, operators and regulators and to promote the international harmonisation of regulatory standards.
One example of preparations for the use of disruptive technologies at international level is the OECD Nuclear Energy Agency's RegLab Project. The aim of the initiative is to test new technologies such as AI, robotics and 3D printing under realistic but controlled conditions. The focus is on so-called sandboxing rounds – structured test phases in which new approaches are simulated and evaluated without having to be used in approved or productive environments. This allows safety aspects, regulatory requirements and practical applications to be systematically analysed. The findings will be used to further develop existing regulations and derive recommendations for national and international committees.
The grid integration of nuclear installations also raises new regulatory issues – especially when AI is used for automated power control. Systems that evaluate technical parameters in real time and derive control impulses are not only required to be functionally reliable but also transparent and eligible for approval in terms of safety. In this context, the Chinese energy authority NEA is focusing on the technical assessment of such systems and their role in the interaction between reactor operation and grid stability.
AI at GRS – research and application
GRS is intensively involved in the assessment and qualification of AI systems for safety-critical applications. As part of a project funded by the Federal Office for the Safety of Nuclear Waste Management (BASE), for example, a modular assessment scheme was developed that takes into account the special characteristics of AI – such as the fact that many AI systems do not operate deterministically. This means that they do not always deliver exactly the same decision or recommendation, even when the input data is identical. This characteristic fundamentally distinguishes AI from classic software and places special demands on transparency, testability and safety. The aim is to examine and further develop the applicability of classic requirements from software development – such as modularity, testability and robustness – to AI systems.
Moreover, GRS analyses how existing regulations (e.g. nuclear safety standard KTA 3501) can be applied to AI applications and where adjustments are needed. The results are incorporated into the further development of assessment criteria and into the advice provided to authorities and ministries.
In addition to assessing external AI systems, GRS also uses artificial intelligence in its own research projects. The aim is to explore the possibilities of data-driven methods for safety-related issues and to develop practical applications. One example is a research project to support the dismantling of nuclear installations funded by the Federal Ministry of Education and Research (BMBF). Here, GRS is working with partners to develop a hybrid learning platform with VR and AR elements and an AI-based system for dismantling-specific knowledge management. The aim is to train specialists more efficiently and make licensing processes more transparent.
Funded by the Federal Environment Ministry, GRS investigated the use of artificial neural networks to generate so-called critical parameter curves, which play a central role in nuclear safety research. Deriving these curves has been computationally intensive up to now. The study showed that neural networks can reproduce real data well across wide ranges – especially in highly non-linear boundary areas, where they significantly outperformed classical interpolation methods.
In a research project commissioned by the Swiss regulatory authority ENSI, GRS colleagues are further developing the so-called adaptive Monte Carlo simulation method to identify critical parameter ranges in the event of loss of coolant accidents in boiling water reactors. Here, machine learning methods are being used to efficiently analyse complex relationships between input parameters and safety-relevant results.
Finally, another example is the work carried out by GRS in collaboration with partner institutions on the development of methods for assessing the remaining service life of metallic components in nuclear power plants. The aim is to better identify and assess ageing processes under realistic operating conditions, such as thermal cyclic loads during load-following operation. Machine learning techniques, in particular neural networks, are also used to identify patterns in measurement data from non-destructive testing methods and to predict fatigue-induced degradation. Here, AI serves as a tool for data-based analysis with the aim of monitoring the ageing process and detecting degradation at an early stage.
In summary, artificial intelligence offers new opportunities for nuclear technology – for example, for optimising technical processes, providing operational support or automating documentation tasks. At the same time, it poses new challenges in terms of safety and regulatory integration. GRS aims to critically monitor these developments, realistically assess their potential and systematically analyse their risks. For one thing is clear: the use of AI in nuclear technology must be transparent, robust and compliant with regulations so that safety is not compromised.