The severe accident scenario propagation studies of nuclear power plants (NPPs) have been one of the most critical factors in deploying nuclear power for decades. During an NPP accident, the accident scenario can change during its propagation from the initiating event to a series of accident sub-scenarios. Hence, having time-wise updated information about the current type of accident sub-scenario can help plant operators mitigate the accident propagation and underlying consequences. In this work, we demonstrate the capability of machine learning (Decision Tree) to help researchers and design engineers in finding distinctive physical insights between four different types of accident scenarios based on the pressure vessel's maximum external surface temperature at a particular time. Although the four accidents we included in this study are considered some of the most extensively studied NPPs accident scenarios for decades, our findings shows that decision tree classification could define remarkable distinct differences between them with reliable statistical confidence.

Supervised machine learning (SML) techniques have been developed since the 1960s. Most of their applications were oriented towards developing models capable of predicting numerical values or categorical output based on a set of input variables (input features). Recently, SML models' interpretability and explainability were extensively studied to have confidence in the models' decisions. In this work, we propose a new deployment method named Decision Tree Insights Analytics (DTIA) that shifts the purpose of using decision tree classification from having a model capable of differentiating the different categorical outputs based on the input features to systematically finding the associations between inputs and outputs. DTIA can reveal interesting areas in the feature space, leading to the development of research questions and the discovery of new associations that might have been overlooked earlier. We applied the method to three case studies: (1) nuclear reactor accident propagation, (2) single-cell RNA sequencing of Niemann-Pick disease type C1 in mice, and (3) bulk RNA sequencing for breast cancer staging in humans. The developed method provided insights into the first two. On the other hand, it showed some of the method's limitations in the third case study. Finally, we presented how the DTIA's insights are more agreeable with the abstract information gain calculations and provide more in-depth information that can help derive more profound physical meaning compared to the random forest's feature importance attribute and K-means clustering for feature ranking.

Nuclear reactor core degradation and relocation happen when the reactor core starts melting and falling into the reactor pressure vessel's lower head during a severe accident. This process encompasses numerous intertwining complex physical phenomena. Researchers and design engineers use integrated severe accident analysis tools such as MELCOR to simulate the core degradation process. However, integrated simulation codes produce tens if not hundreds of thousands of results. Analysing these high-dimensional results is challenging. Currently, these results are being analysed subjectively according to the user. In this work, we present a method for objectifying the analysis of high-dimensional data using explainable machine learning, as represented in our developed tool, the decision tree insights analytics (DTIA) framework. DTIA leverages machine-learning interpretability and explainability to extract insights that correlate input parameters with categorical outputs. We applied DTIA to time-series data for a severe accident in a Nordic boiling-water reactor. DTIA highlighted the different COR variables that are highly associated with different time windows. Analysing these variables led to a more in-depth understanding of the core relocation and degradation. It also suggested some hypothesis that needs confirmation or rejection. Finally, it left some open questions that need further investigation. In summary, DTIA objectively identified correlated variables of high interest and provided insights that were overlooked in previous analyses in the literature. This work presents a step towards objectifying the core analysis of relocation and degradation. However, interpreting the variable insights extracted from DTIA remains subjective, according to the subject-matter expert. 

Nuclear reactor severe accident (SA) progression modelling, computation and simulations are essential for developing severe accident management strategies. Properly developing severe accident management strategies will reduce the risk of radioactive release to the environment. The reactor pressure vessel is the last physical barrier before the core melt leaks into the cavity and containment building. That being said, identifying the mode, timing, and in-depth details of reactor pressure vessel failure under different accident scenarios is essential. Finite element modelling (FEM) is considered the gold standard in structural analysis. However, FEM in general has two main challenges: 1) the huge amount of data associated with the detailed nodal solutions, and 2) the subjectivity in the dependence on visual assessment of different structure variables in comparative studies. Hence, most studies cover the structural analysis globally, by analysing and reporting the significant key variables such as temperatures, stresses and strains. In this work, we tackle those challenges using machine learning interpretability and explainability, leveraging the Decision Tree Insights Analytics (DTIA) tool. DTIA compared the structural analysis variables at the time of RPV failure in station blackout (SBO) and SBO in the presence of a loss of coolant accident (SBO+LOCA). Results lead to the segmentation of the RPV section, highlighting the similarities and differences in the structural variables for the two considered SAs. The results revealed some previously overlooked aspects, including the tensile and compressive behaviour of the reactor pressure vessel’s lower head during various accidents and its effect on the failure time and buckling state before failure.  Finally, the method presented in this article is transferable to perform a comparative study for any comparative structural analysis study.

Due to the complexity of nuclear reactor systems, analysing the high-dimensional data from severe accident (SA) computational tools can be challenging. Although global analysis of SA propagation is standardised to some extent based on regulatory requirements, the detailed analysis remains subjective. This work proposes a framework based on machine learning interpretability and explainability techniques for high-dimensional data analytics. The framework utilises K-means clustering, complemented by a decision tree insights analytics (DTIA) tool, for detailed time-wise analysis of severe accident progression. K-means clustering proposes time windows during which significant events occur, eliminating the subjectivity of the researcher or design engineer. Then, DTIA associates the different parameters with the prespecified time windows. These parameters are then projected into the physical knowledge domain, providing an objective, in-depth understanding of SA propagation.