New AI Framework Boosts Nuclear Safety Confidence

In nuclear engineering, where safety decisions can have profound consequences, assessing the reliability of computer simulations is critical. Researchers from North Carolina State University have developed a new AI-driven framework that makes this process more transparent and consistent, helping to reduce uncertainties in risk analysis. This advancement is vital for ensuring that simulations used in nuclear power plant safety evaluations are trustworthy, potentially preventing costly errors and enhancing public safety.

The key finding of this research is the Predictive Capability Maturity Quantification using Bayesian Network (PCMQBN), which formalizes how to determine if a simulation is adequate for its intended use. Adequacy here means how well the simulation represents real-world quantities of interest under specific scenarios, such as predicting hydrodynamic forces during external flooding events. Unlike previous methods that relied heavily on implicit expert judgments, PCMQBN quantifies this assessment using probabilities, making it easier to understand and apply.

Methodologically, the researchers built PCMQBN on argumentation theory and Bayes' theorem, structuring the decision-making process into a Bayesian network—a graphical model that shows relationships between different pieces of evidence. They characterized evidence from validation activities, such as comparing simulation predictions to experimental data, and assessed factors like data relevance, scaling, and uncertainty. For example, in a case study evaluating Smoothed Particle Hydrodynamics (SPH) for flooding scenarios, they used benchmarks like dam break and moving solid-fluid tests to gather this evidence.

Results from the paper show that PCMQBN improved confidence in simulation adequacy. In the SPH case, when evidence from both benchmarks was integrated, the simulation was deemed 83% adequate, compared to contradictory conclusions from qualitative methods. Sensitivity analysis revealed that validation result dependencies had the highest impact on adequacy assessments. The framework also reduced expected monetary losses in risk analysis by 30%, from about $71.67 in traditional approaches to $50, by providing a more informed basis for decisions.

In practical terms, this framework matters because it helps nuclear engineers and regulators make better-informed safety decisions without relying on vague judgments. For instance, in assessing potential damages from water waves hitting structures, PCMQBN's quantified approach can prevent over- or under-estimation of risks, leading to more cost-effective safety measures. This could extend to other high-stakes fields where simulation reliability is crucial, such as aerospace or environmental modeling.

Limitations noted in the paper include assumptions that the simulation code is verified and that error distributions are simplified for initial development. The method's parameters, like belief scales and conditional probabilities, require refinement as more data becomes available, and it may not fully capture all psychological factors in decision-making under uncertainty.