The views expressed in this post are those of the author and not necessarily those of the Open Nuclear Network or any other agency, institution, or partner.
Artificial intelligence (AI) covers numerous techniques designed to approximate task performance typically associated with human intelligence, from simple preprogrammed chatbots and various statistical tools to artificial neural networks (ANNs). In the nuclear domain, most discussions on AI-related risks remain at the strategic level, such as AI-driven arms races or nuclear escalation, albeit with its speculative nature due to the lack of concrete evidence.[1] AI still has many limitations in meeting such hypes for reasons including its limited capacity to conduct chains of logic and the scarcity of data concerning nuclear system failure.[2]
Meanwhile, AI has converged with nuclear applications as incremental enhancement of functions at the practical level. For instance, a Chinese study conducted in 2017 claims that it succeeded in optimizing centrifuge cascade by using a genetic algorithm (GA) to decrease the total number of centrifuges.[3] In 2024, Russia applied regression trees to predict spent fuel compositions as a function of several inputs, such as initial enrichment levels.[4] In 2021, the United States used an autoencoder for its indirect drive inertial confinement fusion (ICF)[5] program to enhance thermonuclear simulations, which plays a crucial role in understanding the behaviour of nuclear weapons as a part of the Stockpile Stewardship and Management Program.[6][7]
This article explores ongoing AI convergence with the nuclear field at the practical level, focusing on nuclear safety, security, and safeguards. It details how States apply AI techniques to enhance functions, citing scientific publications. The article then examines the potential political, practical, and technical challenges in leveraging the benefits of AI at the international level, especially for verification purposes.
Nuclear Safety
Nuclear safety aims to protect human beings from potential nuclear incidents. For safer and more economically efficient operation of nuclear reactors, it is important to ensure that local power peaks do not exceed safety limits.[8] If a certain area of the nuclear fuel assembly becomes overheated, it narrows the safety margin of the reactor and could reduce the overall power output. AI can contribute to nuclear safety by optimizing fuel loading patterns to flatten power output across the fuel assembly, which contains fuel rods with varying composition and different enrichment levels.[9]
Among AI techniques, GA is a heuristic search method well suited for such optimization tasks. Specifically, GA is useful for identifying near-optimal solutions by simulating the process of genetic evolution. This involves selecting the most promising ones from a set of solutions, then applying crossover and mutation to those selected to produce a new generation of solutions.[10] This iterative process continues until a solution that closely meets the objectives is found.
As examples, recent studies conducted by Iran, North Korea, and China show how GAs could contribute to fuel-loading pattern optimization. In 2013, Iranian researchers used GA in their reactor simulations to determine optimal fuel configurations for the VVER-1000, a 1000 MWe pressurized water reactor (PWR) developed by Russia.[11] In 2021, North Korea adopted an adaptive genetic algorithm (AGA) to enhance the fuel-loading pattern for a PWR, demonstrating AGA's ability to dynamically adjust its crossover and mutation, thereby outperforming static conventional GA methods.[12] Similarly, in 2019, China used AGA for its advanced high temperature reactor (AHTR), a molten salt reactor variant.[13] Even at the simulation stage, these examples highlight recent progress in AI research with the potential of enhancing nuclear safety.
Nuclear Security
Nuclear security aims to prevent theft of nuclear materials or sabotage of facilities by non-State actors or any entities with malicious intent. Nuclear forensics plays a crucial role in nuclear security by using isotopic analysis to identify the potential sources and history of interdicted nuclear material, thereby helping national authorities in their investigation and attribution of the nuclear material to potential sources. Machine learning techniques such as support vector machine (SVM) and K-means clustering are useful in grouping similar isotopic data to enhance forensic analysis. In addition, methods such as multivariate regression could aid in reconstructing material histories. These machine learning tools support nuclear security by clarifying the origins and pathways of nuclear materials.
For instance, in 2022, researchers from Italy and Germany jointly explored the application of SVM for classifying uranium ore concentrates (UOCs) for nuclear forensic purposes. Generally, SVM is used to find decision boundaries that separate data points into distinct classes within multidimensional spaces, comparable to paving a highway that delineates different administrative areas. The study analysed datasets covering multiple features of the concentrates to categorize 79 commercial UOCs into distinct colour groups using SVM.[14] Another example involves studies conducted by a US university, one of which attempts to apply various machine learning techniques, including SVM and K-means clustering, to attribute isotopic ratios of separated plutonium, such as 40Pu/239Pu and 135Cs/137Cs, to one of nine reactor types, including PWRs, MAGNOX, NRX, and CANDU.[15] Given the high degree of accuracy in the test results, the study claims that it successfully demonstrated promising potential of AI for nuclear forensics. These examples underscore the advancing integration of AI in nuclear security.
Nuclear Safeguards
Nuclear safeguards aim to prevent or deter States from illegally developing nuclear weapons or to detect such activities early. The International Atomic Energy Agency (IAEA) applies safeguards in accordance with its safeguards agreements with States. Safeguards activities include nuclear material accounting, containment and surveillance, and inspection. A recent study shows AI’s potential contribution to surveillance activities.
In 2023, United States’ Sandia National Laboratories introduced the Limbo dataset, which combines real and synthetic images of uranium hexafluoride containers. Recognizing a scarcity of publicly available real-world data concerning nuclear safeguards, the study developed the synthetic images on which object detection and image classification models were trained. Then, the trained models were tested on real-world data. The study underscores that computer vision models could learn features irrelevant to target objectives during a training process, which could lead to severe consequences for the non-proliferation community.[16]
Conclusion and Considerations
Before envisioning the over-hyped AI-driven revolutionary transformation in future nuclear power competition, it is crucial to note that incremental functional advancements are made by AI also at the practical level. Many studies have demonstrated that AI-driven optimization techniques, such as GA and AGA, can contribute to the safe and efficient operation of nuclear reactors. Classification and clustering techniques, including SVM and K-means clustering, have been proven useful for nuclear forensics. Lastly, image classification and object detection models could be considered to strengthen nuclear safeguards.
Meanwhile, in harnessing benefits of AI at the international level, these developments pose significant questions for the international community, particularly those concerning future multilateral efforts for safeguards verification activities, which have political, practical, and technical implications, as discussed below.
First, integrating AI-driven surveillance into existing safeguards frameworks may raise political concerns among non-nuclear weapon States (NNWS). This stems from perceptions of the Treaty on Non-proliferation of Nuclear Weapons (NPT) as ‘intrinsically unfair’, allowing only five nuclear weapon States (NWS) to legally hold nuclear weapons.[17] In such a political climate, introducing new tools that enhance verification mechanisms for NNWS' adherence to non-proliferation commitments necessitates thorough discussions with NNWS amid the lack of disarmament progress by NWS, further complicated by the practical and technical hurdles described below.
Second, a practical challenge is posed by the availability of data, which could significantly affect the political acceptability of AI-enhanced surveillance systems. As mentioned, the US study indicates a scarcity of publicly available real-world safeguards information. Consequently, the US research trained its computer vision models on synthetic data, later testing these models with limited real-world data. Without sufficient real-world information that reflects the unique nuclear environments of IAEA Member States, it is uncertain that future AI-driven surveillance system could exhibit reliable performance across all NNWS. More importantly, the capability of AI models could experience significant decline when excessively dependent on synthetic data in training processes.[18],[19] Given the uncertainty in capabilities of AI systems trained on synthetic data, garnering political acceptability by NNWS would be challenging.
Third, setting performance standards for AI surveillance in nuclear safeguards presents a technical challenge. The evaluation of AI effectiveness involves multiple metrics that often have inverse relationships. In other words, creating AI models that excel in all metrics simultaneously is a challenging task, and developers must thus prioritize certain metrics over others depending on a system's objectives, exemplified as precision-recall trade-off. Aiming to capture all potential illicit diversion activities could lead to the inclusion of some false positives, causing false alarms for compliant situations. Conversely, prioritizing higher confidence in detected diversion activities could lead to missing some actual proliferation cases. This balance is crucial in nuclear safeguards, where the potential repercussions of AI-induced false alarms or overlooked proliferation signs have significant political and proliferation implications. Furthermore, also the question of who decides AI performance standards would be perplexing.
To navigate the complexities of political acceptability, data availability, and technical standards, it is advisable to foster inclusive development of AI. This entails engaging NWS and NNWS in cooperative research initiatives, data-sharing agreements, and standard-setting discussions under the auspices of international bodies such as the IAEA. Such collaborative efforts should prioritize transparency, equality, and adherence to global non-proliferation and disarmament objectives, ensuring that AI applications in nuclear domains serve the collective interests of peace and security.
[1] Vincent Boulanin, et al., Artificial Intelligence, Strategic Stability and Nuclear Risk, Stockholm International Peace Research Institute, 2020, https://www.sipri.org/publications/2020/policy-reports/artificial-intelligence-strategic-stability-and-nuclear-risk.
[2] Cheng, B.; Bradley, P.A. What Machine Learning Can and Cannot Do for Inertial Confinement Fusion. Plasma 2023, 6, 334–344. https://doi.org/10.3390/plasma6020023
[3] Optimization of Centrifuge Purge Cascade with Additional Feed Flow, Yuanzineng Kexue Jishu/Atomic Energy Science and Technology, Vol.51, Issue 9, September 2017. https://doi.org/10.7538/yzk.2016.youxian.0781
[4] Md. Tarequzzaman and Alexander Nakhabov, "Prediction of Spent Nuclear Fuel Isotopic Composition for the VVER-1000 Reactor Utilizing Regression Tree," Annals of Nuclear Energy 195 (January 2024): 110161, https://doi.org/10.1016/j.anucene.2023.110161.
[5] Hyuk Kim, "North Korea’s Nuclear Fusion Research," 38 North, February 2022, https://www.38north.org/2022/02/north-koreas-nuclear-fusion-research/.
[6] K. D. Humbird, J. L. Peterson, J. Salmonson, and B. K. Spears, Cognitive simulation models for inertial confinement fusion: Combining simulation and experimental data, Physics of Plasmas, Vol.28, Issue 4, April 2021. DOI: https://doi.org/10.1063/5.0041907
[7] Joseph D. Morelle, Rick Larsen, Eric Swalwell, Zoe Lofgren, Bill Foster, Scott H. Peters, Jim Costa, Mark Takano, Raja Krishnamoorthi, and Abigail Davis Spanberger, Jimmy Panetta, "Letter to The Honorable Mike Simpson and The Honorable Marcy Kaptur Regarding FY24 Inertial Confinement Fusion (ICF) Program," March 31, 2023, Congress of the United States, House of Representatives, Washington, DC 20515. Available at: https://lofgren.house.gov/sites/evo-subsites/lofgren-evo.house.gov/files/FY24%20Inertial%20Confinement%20Fusion%20%28ICF%29%20Program.pdf
[8] "Hot Channel Factors - Peaking Factors," Nuclear Power. Available at: https://www.nuclear-power.com/nuclear-power/reactor-physics/reactor-operation/normal-operation-reactor-control/hot-channel-factors-peaking-factors/.
[9] Hyuk Kim, "North Korea’s Artificial Intelligence Research: Trends and Potential Civilian and Military Applications," 38 North, January 2024, https://www.38north.org/2024/01/north-koreas-artificial-intelligence-research-trends-and-potential-civilian-and-military-applications/.
[10] Ibid
[11] Rafiei Karahroudi M., Mousavi Shirazi S.A., and Sepanloo K., “Optimization of designing the core fuel loading pattern in a VVER-1000 nuclear power reactor using the genetic algorithm”, Annals of Nuclear Energy, Vol.57, 2013. https://doi.org/10.1016/j.anucene.2013.01.051