As the push for decarbonization intensifies, the need for innovative and efficient technologies in the energy sector has never been more urgent. Among the emerging technologies, nuclear energy is positioning itself as a pivotal player in reducing carbon emissions. The ongoing development of new nuclear reactors and the optimization of existing ones are crucial to meeting global decarbonization goals. While machine learning (ML) and artificial intelligence (AI) offer promising advancements for enhancing reactor design and improving safety measures, these technologies face significant challenges in aligning with the rigorous safety regulations set by the U.S. Nuclear Regulatory Commission (NRC). The regulatory hurdles posed by the NRC could slow the adoption of AI, particularly in areas critical to nuclear safety.
The Regulatory Conundrum of AI in Nuclear Safety
The involvement of AI and ML in the nuclear energy sector represents a groundbreaking shift towards more efficient and automated processes. AI models offer unparalleled speed in identifying patterns, which can accelerate the design of nuclear reactors and enhance the monitoring of safety-critical parameters in real time. Machine learning algorithms can, for instance, be used to predict how a reactor design might behave under various conditions or anticipate equipment failures before they occur, potentially reducing risks and improving reactor longevity.
However, this rapid adoption of AI in nuclear energy comes with a significant challenge: model transparency. To comply with NRC standards, any technology used in nuclear applications, including AI models, must be explainable. Simply put, if an AI algorithm is used to predict a safety threshold for a reactor’s operation or assess its safety features, the NRC must be able to validate the algorithm’s reasoning and understand how it arrived at its conclusions.
The issue here is that many AI models function as “black boxes,” meaning that the pathways connecting their inputs to outputs are often opaque. This inherent lack of transparency is a major obstacle for traditional regulatory procedures. Nuclear regulators rely on well-established processes and methods to ensure safety and reliability, and the opacity of AI systems makes it extremely difficult to apply these processes in an effective manner. For a safety-sensitive field like nuclear energy, where any misstep could have catastrophic consequences, this challenge becomes even more critical.
In response to this regulatory bottleneck, there is a pressing need for tools that can bridge the gap between the promise of AI technology and the rigorous scrutiny required by regulators. This is where explainable AI (XAI) for nuclear energy comes into play, and researchers at the University of Michigan are at the forefront of developing such solutions.
The Development of pyMAISE: A Step Toward Licensing AI for Nuclear Safety
The University of Michigan research team, led by assistant professor Majdi Radaideh, has begun work on a groundbreaking project called pyMAISE (Python-based Michigan Artificial Intelligence Standard Environment). pyMAISE is an innovative benchmarking library designed specifically for the nuclear industry, with the aim of making machine learning models in the field more explainable and easier for regulators like the NRC to assess. By creating this tool, the researchers hope to provide nuclear engineers and regulators with a common platform that will allow them to quickly evaluate the potential of AI in nuclear energy applications while addressing safety concerns.
Radaideh explains that pyMAISE’s primary goal is to create a standardized method for testing explainable AI in nuclear engineering applications. “PyMAISE is one step to help the NRC create a pipeline for licensable AI,” he states. The software package aims to simplify the process of developing and testing machine learning models, allowing nuclear engineers—many of whom may not have a deep background in AI—to create tools that they can apply to safety assessments and reactor design simulations with minimal effort. This democratization of AI is crucial to accelerating the pace of innovation while ensuring that all tools are accountable and understandable.
Key Features of pyMAISE: Empowering Engineers with Simplified AI and Safety Modeling
At its core, pyMAISE helps nuclear engineers pinpoint the best machine learning models for a wide range of applications, from basic linear regressions to more complex neural networks. The library allows for rapid experimentation and fine-tuning by automatically testing a broad variety of models, assessing their performance across different datasets, and evaluating their predictive accuracy.
In nuclear reactor design, for instance, pyMAISE can model how various design parameters influence reactor power output. The software leverages simulated data and machine learning to identify optimal parameters, potentially leading to the creation of more efficient and cost-effective microreactors and other advanced nuclear designs.
Safety monitoring is another critical area in which pyMAISE proves valuable. In one of the study’s use cases, the tool was used to model and predict the critical heat flux (CHF) of a reactor—a safety-critical parameter that determines the maximum heat a reactor can sustain before the reactor core overheats and becomes unstable. Given the complexity of this calculation, AI and ML techniques can help quickly identify potential safety hazards.
PyMAISE has already shown promising results in various applications, performing on par or better than well-established ML benchmarking tools such as Auto-Sklearn, AutoKeras, and H2O.ai. In these tests, pyMAISE often explored more potential models, sometimes making use of fewer training samples—demonstrating its ability to work with limited data and provide effective results. For instance, in the reactor design case, pyMAISE quickly zeroed in on the ideal model for predicting reactor power output, without requiring extensive datasets.
Building Transparency into Machine Learning Models: A Crucial Step for Regulatory Approval
One of the most groundbreaking features of pyMAISE is its inclusion of explainability features, a rare but increasingly sought-after tool in the AI world. The tool can identify the most important features in a model and reveal how those features influence the outcomes. For nuclear energy, where a deeper understanding of model predictions can be the difference between safe operations and catastrophic failures, this level of transparency is a game-changer.
While many machine learning algorithms operate as “black boxes,” making it nearly impossible for human experts to interpret their reasoning, pyMAISE allows the operator to discern why a particular input generated a certain output. The team’s hope is that this transparency will not only facilitate the evaluation and approval process by the NRC but also empower nuclear engineers and regulators to make safer, more informed decisions when considering the application of machine learning in reactor design and operational safety.
Potential Applications and Impact Beyond Nuclear Energy
While pyMAISE is initially aimed at the nuclear industry, the potential applications of explainable AI reach far beyond this field. Any industry where AI is deployed for safety-critical tasks—such as healthcare, finance, or aerospace—would benefit from more transparent and interpretable models. As AI continues to be integrated into high-stakes sectors, ensuring that these models can be understood, explained, and validated by human experts becomes an urgent priority.
In healthcare, for example, AI models are being developed to predict patient outcomes, detect diseases, and guide treatment plans. In finance, algorithms are being used to predict market trends or assess the risk of loans. In all these cases, stakeholders—whether they are doctors, investors, or regulators—need to be able to trust and understand the model’s outputs. PyMAISE’s focus on explainability could serve as a foundation for similar efforts in these industries, promoting widespread adoption of AI technologies without compromising safety.
Looking Ahead: The Future of AI in Nuclear Energy
As nuclear energy advances toward its role in global decarbonization efforts, pyMAISE represents a critical step in integrating AI into the nuclear industry without compromising on safety. As the NRC and other regulatory bodies start embracing machine learning tools, pyMAISE’s explainability features will ensure that the benefits of AI can be realized without undermining public trust in the safety and reliability of nuclear power.
This innovation could transform how nuclear reactor designs are approached and managed, streamline the process of safety evaluation, and speed up advancements in reactor technology. But perhaps most importantly, it represents an essential evolution of regulatory practices, allowing safety guidelines and new technological capabilities to coexist, fostering a safer, more efficient nuclear industry in the process.
The future of nuclear energy is not only about harnessing powerful new technologies like AI but doing so in a way that maintains the highest standards of safety and accountability. Through tools like pyMAISE, this balance becomes more achievable, ensuring that AI can be leveraged for innovation while staying true to the foundational principles of safety and transparency that the nuclear industry demands.
Reference: Patrick A. Myers et al, pyMAISE: A Python platform for automatic machine learning and accelerated development for nuclear power applications, Progress in Nuclear Energy (2024). DOI: 10.1016/j.pnucene.2024.105568