Students
Forrest Shriver, Aidan Furlong
Summary
The use of machine learning algorithms and neural networks to model human behavior, predict physics parameters, and to mine data is an ever-growing field of research. We are at the forefront of the Artificial Intelligence (AI) solutions and their application to a wide array of problems. They can be used to explore new design spaces that have never been investigated, used to discover new features from data, and accelerate computationally expensive simulations.
To date, AI has only played a small role in the analysis of nuclear systems; however, this is starting to change. The use of well-trained neural networks as surrogate models for complex systems is starting to be used as reduced order models (ROMs). Nuclear reactors are complex systems that require detailed knowledge of different physics problems. Historically, the complexity of nuclear systems, has led to the necessity of splitting the analysis into different single physics software packages. This is referred to as operator splitting. As computational resources continue to expand, the size and accuracy of the computational models improve. The primary method for further improving analysis is to couple separate physics codes together to preform analysis with more complexity that include feedback terms that were not possible before. Even with these advancements, due to the size and complexity of the systems being analyzed, a coarse volume averaged representation (nodalization) of the fine structures of the system is still required to perform optimization, design, and safety analysis. This coarse nodalization leads to inaccuracies in localized parameters that are important for nuclear reactor safety margin prediction, i.e. peak pin power. These inaccuracies can be improved using high-speed high-fidelity methods. Currently, high-fidelity results require a significant amount of computational time and are not useful for performing repeated analysis like parametric studies. This is where data driven approaches called surrogate models can make large impacts accelerating product design, optimization, analysis, and monitoring. In this work we seek to develop a hierarchical, flexible, integrate neural network (HIFI-NN) as a surrogate model that will be capable of inferring nuclear reactor core parameters in milliseconds. The methods developed will be applicable beyond nuclear engineering and suitable for computer vision applications with complex parameter interdependence such as transportation, healthcare, and manufacturing. The demonstration model chosen for this project is a PWR reactor core due to the abundance of data and verified and validated software packages available. However, the HIFI-NN will be readily extendable to other reactor designs including GenIV. For example, modeling the neutronics and thermal-hydraulics of the Generic Pebble-Bed Fluoride-salt-cooled High-temperature Reactor (gFHR) where full core Monte Carlo cross sections are used.
Researchers have been searching for methods to accurately perform reactor core survey calculations and optimization studies using machine learning techniques for many years, with various degrees of success. The majority of the research up to this point focused on the prediction of assembly average parameters and core average parameters for burnup calculations, or loading pattern (LP) optimizations.
Preliminary work has been completed with the development of LatticeNet. The methodology LatticeNet was based on combines computer vision approaches, modular neural network approaches, and hyperparameter optimization methods to intelligently design novel network architectures. This methodology has been used to develop a architecture which is capable of predicting pin-resolved powers and keff at a high level of detail. The results produced show the successful prediction of the target neutronics parameters under a variety of typical neutronics conditions, and they indicate a potential path forward for neural network–based model development. Its ability to predict pin powers and other parameters within single or multiple 2D pressurized water reactor assemblies was demonstrated. It has been shown to be effective at predicting distributions of reactor parameters such as normalized pin powers under changing thermal hydraulics conditions. However, scaling this architecture to problems beyond a mini-core may be difficult.
Based on the experience with LatticeNet we are looking to construct a new AI architecture, HIFI-NN, that also trains on edge and volume averaged parameters. In this way single assembly networks can be trained and assembled to form a full core representation without modeling the full core geometry. therefore we are focusing on high-fidelity parameter prediction on the sub-assembly level that will be used to accelerate transport solutions or increase the fidelity of coarse volume averaged solvers. We will demonstrate the ability of the network to accelerate transport solutions and predict relevant high-fidelity neutronics parameters for a full PWR reactor core. This will be accomplished by using the HIFI-NN as a surrogate model for higher-fidelity solutions. This surrogate model will be coupled to the NRC’s TRACE/PARCS system analysis package and verified against standard benchmark problems. The new hybrid system will have a clear computational advantage over existing state-of-the-art coupled system analysis codes.
Publications
Furlong, A., Watson, J. K., Shriver, F., “Investigation of Monte Carlo trained CNNs for neutronics predictions of typical and atypical PWR assemblies,” Progress in Nuclear Energy, Vol. 166, 104961, November (2023)
Furlong, A., Shriver, F., Watson, J. K., “Using Neural Networks to Predict Pin Powers in Reflective PWR Fuel Assemblies with Varying Pin Size,” PHYSOR 2022, Pittsburgh Pa, May 15-20, (2022)
Shriver, F., Watson, J. K., “Physically Adversarial Thermal Hydraulics Evaluation of Deep Learning Models for Pressurized Water Reactors,” Progress in Nuclear Energy, doi:10.1016/j.pnucene.2022.104149 (2022)
Shriver, F., Watson, J. K., “Scaling Deep Learning for Whole-Core Reactor Simulation,” Progress in Nuclear Energy, doi:10.1016/j.pnucene.2022.104134 (2022)
Shriver, F., Gentry, C., Watson, J. K., “Prediction of Neutronics Parameters within a 2D Reflective PWR Assembly Using Deep Learning,” Nuclear Science and Engineering, Taylor & Francis, 0, 1-22 (2021)
Patents
Provisional Patent Applications:
Shriver, F., Watson, J. K., Gentry, C., “Methods for Prediction of Neutronics Parameters Using Deep Learning,” T18396 (222107-8575), U.S. Provisional Patent Application Serial No. 63/123,260, filed December 9, 2020