Nuclear Energy Agency (NEA) – Task Force on Artificial Intelligence and Machine Learning for Scientific Computing in Nuclear Engineering: Purdue University Reactor 1 (PUR-1) Exercise

Recent advances in artificial intelligence (AI) and machine learning (ML) techniques have increased interest in using these computational tools for reactor condition monitoring, signal analysis, anomaly detection, and condition-based maintenance. Increased digitization of sensors in both current and future builds will provide unprecedented real-time information on all aspects of plant operations. You can get a rich and detailed understanding of the condition of your plants. While it is difficult to fully analyze vast amounts of data using traditional methods, new methods in data analytics, machine learning, and AI can be used to extract insights and support decision-making.

However, nuclear safety regulatory requirements differ from typical AI/ML applications. Critical gaps include task-specific change requirements as well as AI/ML validation, validation, uncertainty quantification, data deficiencies, and model interpretability. To address these gaps, the NEA Task Force on Artificial Intelligence and Machine Learning for Scientific Computing in Nuclear Engineering was established within the Expert Group on Reactor System Multiphysics (EGMUP) of the Working Group on Scientific Problems and Uncertainty Analysis of Nuclear Systems (WPRS) to design benchmark exercises across a variety of computational areas of interest to address some of these challenges and develop AI/ML We were able to develop guidelines for applying the method to nuclear engineering applications.

This benchmark exercise presents a feasibility study for characterizing two types of outages using data from the Purdue University Reactor No. 1 (PUR-1) research reactor at Purdue University in West Lafayette, Indiana, USA. PUR-1 first reached critical status in 1962 and, after reauthorization in 2016, became the first all-digital reactor system in the United States. PUR-1 can collect over 2000 different signals with a time resolution of 1 second, including measurements such as neutron flux, rod position, radiation level, pool temperature, power, and calculated signals such as system rate of change. A series of measurements from PUR-1 was collected, evaluated, and stored in the NEA data bank as the so-called Purdue University Reactor Shutdown Event Database (PURSE), which is redistributed to other NEA organizations. The PURSE dataset consists of real reactor data, making it ideal for benchmarking AI/ML applications. In the future, additional PUR-1 data may enable testing of different reactor conditions and transients that are not possible at operating commercial nuclear power plants.

The benchmark exercise includes three tasks based on time series signals collected from the PUR-1 reactor. The goal of Task 1 is to classify bottom gang and SCRAM shutdown conditions based on selected signals collected over 800 seconds. Task 2 requires classifying shutdown conditions based on truncated signals. In this case, the objective is to determine how soon after a shutdown the type of shutdown initiation can be detected. These tasks evaluate the classification quality, such as accuracy and false positive rate, of the applied algorithms, which are of paramount importance for state-of-the-art predictive maintenance applications. Finally, Task 3 continues to focus on time series forecasting based on the shutdown example. The objective is to predict the future value of a selected signal over time.

The benchmark exercise is supervised by: Task Force on Artificial Intelligence and Machine Learning for Scientific Computing in Nuclear Engineering. Results will be reported to the task force and presented during the annual WPRS Benchmarking Workshop.

Coordinators: William Stephen RICHARDS, Stylianos CHATZIDAKIS (Purdue University, USA), Catalina ANGHEL, Kamal MORAVEJ (Canadian Atomic Energy Research Institute, Canada), Gregory DELIPEI, Xu WU (North Carolina State University, USA)

NEA Secretariat: Oliver Bass

NEA GitLab work area

Participation

All NEA member countries are eligible to participate. Participants are asked to:

1. License PURSE data,
2. To participate in this benchmarking activity, please submit a signed terms and conditions form to the NEA Office.

schedule