Background

Due to recent advances in artificial intelligence (AI) and machine learning (ML) methodologies, there is an increasing interest in using these computational tools for reactor health monitoring, signal analysis, anomaly detection, and condition-based maintenance. The increasing digitisation of sensors in both current and future builds provides unprecedented real-time information about all aspects of plant operation. It gives a rich and detailed picture of the plant status.  While vast amounts of data are difficult to fully analyse using traditional methods, new methods in data analytics, machine learning, and AI can be used to extract insights and support decision-making.

However, the requirements for nuclear safety regulations are different than typical AI/ML applications. Critical gaps include the requirement of task-specific modifications, as well as verification, validation and uncertainty quantification of AI/ML, data scarcity and interpretability of models. 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 Systems Multiphysics (EGMUP) of the Working Party on Scientific Issues and Uncertainty Analysis of Reactor Systems (WPRS) to design benchmark exercises spanning various computational domains of interest, and which could address some of these challenges as well as develop guidelines for applying AI/ML methodologies to nuclear engineering applications.

Scope and objectives

This benchmark exercise presents a feasibility study on the characterisation of two types of shutdowns using data from the Purdue University Reactor Number One (PUR-1) research reactor located at Purdue University in West Lafayette, Indiana, United States. PUR-1 first went critical in 1962, and after re-licensing in 2016, it became the first all-digital reactor system in the United States. The PUR-1 can collect over 2000 different signals at one second time resolution, including measurements such as neutron flux, rod positions, radiation levels, pool temperature, and power, as well as calculated signals such as system change rate. A set of measurements from PUR-1 was collected, evaluated, and archived as so-called “Purdue University Reactor Shutdown Event Database (PURSE)” at the NEA Data Bank, which is redistributing the database to other NEA organisations.  The PURSE dataset is well suited for benchmarking AI/ML applications because it consists of real reactor data. In the future, additional PUR-1 data may allow for the testing of different reactor conditions and transients that would be impossible in an operating commercial nuclear power plant.

The benchmark exercise includes three tasks based on time-series signals collected from the PUR-1 reactor: The goal in Task 1 is the classification of gang lower and SCRAM shutdown states based on selected signals collected over 800 seconds. Task 2 requires the classification of the shutdown state based on a truncated signal; in this case, the aim is to determine how soon after shutdown the type of shutdown initiation can be detected. These tasks will assess the classification quality including precision and the false-positive rate of the applied algorithms which are of utmost importance for state-of-the-art predictive maintenance applications. Finally, Task 3 focuses on time series forecasting, still based on shutdown examples; the aim is to predict the future values over time of selected signals. 

Organisation

The benchmark exercises are supervised by the Task Force on Artificial Intelligence and Machine Learning for Scientific Computing in Nuclear Engineering. Results are reported to the Task Force and will be presented during the annual WPRS Benchmarks Workshops.

Co-ordinators: William Stephen RICHARDS, Stylianos CHATZIDAKIS (Purdue University, United States), Catalina ANGHEL, Kamal MORAVEJ (Canadian Nuclear Laboratories, Canada), Gregory DELIPEI, Xu WU (North Carolina State University, United States)

NEA Secretariat: Oliver BUSS

NEA GitLab Working Area

Participation

Participation is open to all NEA member countries. Participants are asked to

1. obtain a license for the PURSE data and
2. send a signed version of the conditions form to the NEA Secretariat to join this benchmark activity.

Schedule