Artificial Intelligence in Reactor Physics: Current Status and Future Prospects

TL;DR

The paper surveys how AI/ML, particularly physics-informed neural networks (PINN) and related variants, are applied across reactor physics to solve neutron transport/diffusion equations, predict key state parameters, and enable real-time monitoring and safety design. It covers steady-state, transient, and burnup problems, highlighting surrogate modeling, digital twins, and data-assimilation techniques as core themes, while noting fragmentation and limited generalization in current work. The authors synthesize methodological trends (PINN families, DeepONet, TL-PINN, cPINN, BD-PINN, DEPINN) and industrial applications, emphasizing multiphysics coupling with thermo-hydraulics and the value of data-driven corrections. They also outline future directions, including theoretical breakthroughs, improved implementation, and robust, explainable models for safe, economical reactor operation.

Abstract

Reactor physics is the study of neutron properties, focusing on using models to examine the interactions between neutrons and materials in nuclear reactors. Artificial intelligence (AI) has made significant contributions to reactor physics, e.g., in operational simulations, safety design, real-time monitoring, core management and maintenance. This paper presents a comprehensive review of AI approaches in reactor physics, especially considering the category of Machine Learning (ML), with the aim of describing the application scenarios, frontier topics, unsolved challenges and future research directions. From equation solving and state parameter prediction to nuclear industry applications, this paper provides a step-by-step overview of ML methods applied to steady-state, transient and combustion problems. Most literature works achieve industry-demanded models by enhancing the efficiency of deterministic methods or correcting uncertainty methods, which leads to successful applications. However, research on ML methods in reactor physics is somewhat fragmented, and the ability to generalize models needs to be strengthened. Progress is still possible, especially in addressing theoretical challenges and enhancing industrial applications such as building surrogate models and digital twins.

Figures (6)

• Figure 1: The process of a fission reaction in a nuclear reactor. A neutron interacts with a heavy nucleus (such as $\mathrm{^{235}U}$ or $\mathrm{^{239}Pu}$), triggering a fission reaction that splits the heavy nucleus into two or more smaller nuclei. This process releases a large amount of energy, mainly as the kinetic energy of the fission fragments, which is then converted into heat. During fission, additional neutrons are emitted, which can be absorbed, scattered, or trigger further fission events. If each fission-produced neutron successfully induces at least one new fission, the reaction becomes self-sustaining, forming a chain reaction. Control mechanisms within the reactor regulate this process by absorbing excess neutrons to maintain stability. The application of ML methods to reactor physics mainly considers the optimization of the parameters of this process, which are derived from the requirements of nuclear power plants.
• Figure 2: Literature Publication Trends in Reactor Physics: This histogram illustrates the annual publication count from the earliest available data through 2023, highlighting a notable surge in research output since 2020. The horizontal axis denotes the years, and the vertical axis shows the number of publications, offering insights into the growing academic interest in this field.
• Figure 3: Application of AI/ML Methods in Reactor Physics: A Comparative Analysis. This bar chart categorizes and compares the frequency of various AI/ML methodologies in the literature, including CNN, DNN, RNN, Reinforcement Learnin (RL), Transfer Learning (TL), Long and Short-Term Memor (LSTM), and Fully Convolutional Networks (FCN) within Deep Learning (DL), as well as the specialized PINN bib:49 approach. The x-axis lists the ML methods, and the y-axis indicates the corresponding number of publications, reflecting the diverse algorithmic applications in reactor physics research.
• Figure 4: In this paper, we delineate the taxonomy and interplay of ML methods. Traditional methods, particularly those grounded in ensemble learning, frequently serve as the foundation for predicting core parameters in reactor physics research. Deep learning stands out as the most extensively studied and cutting-edge approach, with a multitude of methodological variants emerging to address specific challenges in the field. Reinforcement learning and transfer learning, as burgeoning disciplines, are poised for further exploration within the domain of reactor physics. These areas hold promise for enhancing the efficiency of learning from data and for optimizing model transferability to mitigate the time costs associated with training. Future research is likely to delve into these methods, investigating their potential to streamline the predictive capabilities and computational efficiency of reactor physics models.
• Figure 5: The mind map shows practical challenges and application scenarios of ML methods in the field of reactor physics. On one hand, this paper takes a cut through some of the challenges that need to be solved urgently in industry, including the steady state and transient problems derived from the neutron spacetime variables, and the burnup problems containing the fuel where the nuclei of the atoms to be reacted are located, which have been facilitated by ML to achieve good effects; on the other hand, there are several industrial applications based on these problems, including advanced data processing methods, and the practical impetus brought by ML in the operational simulation, online monitoring, and safety design.