This report presents the project Idaho National Laboratory conducted for the Nuclear Regulatory Commission (NRC) to explore the advanced computational tools and techniques, such as artificial intelligence (AI) and machine learning (ML), for operating nuclear plants. The report reviews the nuclear data sources, with the focus on operating experience data, that could be applied by advanced computational tools and techniques. Plant-specific and generic (national and international) data from different sources are described. The report describes the relationships between statistics and AI/ML and then introduces the most widely used AI/ML algorithms in both supervised and unsupervised learning. The report reviews the recent applications of advanced computational tools and techniques in various fields of nuclear industry, such as reactor system design and analysis, plant operation and maintenance, and nuclear safety and risk analysis. The report presents the insights from the project on the potential applicability of AI/ML techniques in improving advanced computational capabilities, how the advanced tools and techniques could contribute to the understanding of safety and risk, and what information would be needed to provide meaningful insights to decision makers.The report also documents an NRC survey on the current state of commercial nuclear power operations relative to the use of AI and ML tools as well as the role of AI/ML tools in nuclear power operations, which was published by the NRC as in the Federal Register Notice NRC-2021-0048 in April 2021. A summary of the survey, including the survey questions, survey participants, survey responses, and the conclusions and insights derived from the survey, is provided in the report.Finally, the report investigates potential applications of using AI/ML in operating nuclear power plants and advanced reactors (both advanced light-water reactors and advanced non-light-water reactors) to improve nuclear plant safety and efficiency. Three main application fields are considered: plant safety and security assessments; plant degradation modeling, fault and accident diagnosis and prognosis; and plant operation and maintenance efficiency improvement. v TABLE OF CONTENTS ABSTRACT .

This report presents the initial application design of the Cost Risk Analysis Framework Tool (CRAFT). At the beginning, the concept of CRAFT was formulated from a need to perform economic risk analysis that can be quickly deployable through the nuclear industry to evaluate the risk associated to capital projects and to perform plant asset management. The design evolved into a framework that can integrate several forms of risk (e.g., not just economic, but also safety related) into a single analysis. In order to reduce the research, development and deployment time, we have decided to use the Risk Analysis Virtual ENvironment (RAVEN) statistical framework as a basis for the CRAFT architecture. In this report we summarize the most recent developments accomplished during the second part of FY18 for the two use cases of the Cost and Risk Categorization Applications path of the RISA project. These two use cases directly target plant health and plant capital SSC management. Regarding the plant health management use case, we have started to investigate methods to extract information from text-based data: text data mining. The Risk Informed Asset management is the main target of this report and several methods are here shown in order to solve the capital SSC replacement issue. This has been performed by presenting: several models designed to determine the effective cost through the lifetime of a capital SSCs, and the optimization algorithms that will be employed to determine the optimal replacement schedule of a given set of capital SSC. The interface between RAVEN and the PRA code SAPHIRE is the first step toward the integration of several plant risk models (e.g., economic and safety) and plant data (e.g., databases from the plant maintenance and diagnostics center) into a single analysis framework. In this respect, RAVEN is the ideal platform to connect several models and to manage data streaming among them.

This study examined the relationship between how visual information is organized and people’s visual search performance. Specifically, we systematically varied how visual search information was organized (from well-organized to disorganized), and then asked participants to perform a visual search task involving finding and identifying a number of visual targets within the field of visual non-targets. We hypothesized that the visual search task would be easier when the information was well-organized versus when it was disorganized. We further speculated that visual search performance would be mediated by cognitive workload, and that the results could be generally described by the well-established speed-accuracy tradeoff phenomenon. This paper presents the details of the study we designed and our results.

The current aging management plans of passive structures in nuclear power plants (NPPs) are based on preventative maintenance strategies. These strategies involve periodic, manual inspection of passive structures using nondestructive examination (NDE) techniques. This manual approach is prone to errors and contributes to high operation and maintenance costs, making it cost prohibitive. To address these concerns, a transition from the current preventive maintenance strategy to a condition-based maintenance strategy is needed. The research presented in this paper develops a condition-based maintenance capability to detect corrosion in secondary piping structures in NPPs. To achieve this, a data-driven methodology is developed and validated for detecting surrogate corrosion processes in piping structures. A scaled-down experimental test bed is developed to evaluate the corrosion process in secondary piping in NPPs. The experimental test bed is instrumented with tri-axial accelerometers. The data collected under different operating conditions is processed using the Hilbert-Huang Transformation. Distributional features of phase information among the accelerometers were used as features in support vector machine (SVM) and least absolute shrinkage and selection operator (LASSO) logistic regression methodologies to detect changes in the pipe condition from its baseline state. SVM classification accuracy averaged 99% for all models. LASSO classification accuracy averaged 99% for all models using the accelerometer data from the X-direction.

No abstract