Deep Neural Networks as SurrogateModels for Fuel Performance Codes

Sammanfattning: The core component of a nuclear power plant is the reactor and the fuel rods that supply it with fission fuel. Efficient and safe energy extraction is thus highly dependent on the reactor design and the conditions of the fuel rods. To anticipate high-quality operation and potential risks in advance, one must perform simulations on the fuel rods. This is traditionally executed using fuel performance codes such as the Transuranus and FRAPCON. These codes offer intricate and accurate models of the underlying physical processes that govern fuel rod performance, encompassing aspects like linear heat generation rates, neutron flux and their irradiation effects, fuel rod expansion and contraction, clad corrosion, and fission gas release. However, fuel performance codes cannot be easily parallelized using modern hardware accelerators, such as graphic processing units, due to their iterative nature that follows from the complexity of coupling multiphysics-based models. Thus, they can only simulate one fuel rod at a time per CPU. The challenge comes when trying to simulate all the 10,000 to 100,000 fuel rods in a typical pressurized water reactor, which is a relatively slow process as compared to parallelized computation for all time steps using a GPU. To improve the speed aspect of the fuel performance modeling, a data-driven surrogate model based on neural networks is developed. The main advantage of a neural network over fuel performance codes is its ability to perform parallelized computations using one or more GPUs. The preexisting architectures that were explored include Temporal Convolutional Network, Fourier Neural Operator, and a Transformer. Additionally, a novel architecture is proposed, the Temporal Frequency Network. This is a heterogeneous ensemble method that is based on the Temporal Convolutional Network and the Fourier Neural Operator. The newly proposed architecture archives the lowest validation error among the preexisting architectures with a minor increase in the computations as compared to the ensemble components. The Temporal Frequency Network is then applied to take time-dependent inputs in the form of linear heat generation rate and use it as the only information to predict various time-dependent variables of the fuel rods. The predicted variables include fuel center-line temperature, central void pressure, oxidation thickness, fuel gap width, hydrogen absorption, integral fission gas release, and integral fractional gas release. When deploying the neural network in practice, the user cannot in general trust that the model will generalize from its training, especially in fuel performance modeling where accurate predictions are important to demonstrate safe operation. To ensure that the predictions of the model are reliable, a separate neural network called decoder is trained to qualitatively quantify the error of the previous model that made the predictions, called encoder. This is done by training the decoder to reconstruct the original input to the encoder by providing it with the output, e.g., the inverse task. It is then possible to compare the original input with the reconstructed input, thus, an error can be computed that can be used to qualitatively determine the quality of the predictions.  With the Temporal Frequency Network, the average validation error was roughly 1% error. This makes it a strong candidate surrogate model for fuel performance modeling. In addition, with the encoder-decoder setup, this work provides a robust framework for error estimation that can be performed on new, potentially out-of-distribution inputs without the need for a fuel performance code. This setup is therefore suitable for applications where a low false negative rate is desired.