Conference Agenda

Digital twins are becoming essential in advanced nuclear reactor technology for real-time monitoring, predictive maintenance, and optimization. They help maximize uptime, predict failures, and enable cost-effective testing. However, most digital twins rely on deterministic models that fail to capture uncertainties in sensor data and modeling. While Bayesian models can quantify uncertainty, they require significant computational resources and do not scale well, making them impractical for real-time applications. This work presents the application of variational Bayesian last layer (VBLL) augmented recurrent neural networks (RNNs) to produce uncertainty-aware models while mitigating the issue of computational cost through variational inference. Variational inference allows the neural network to assign parameters to probability distributions rather than point values, and thus permits sampling of predictions to measure the uncertainty of the model. We apply both deterministic Long Short-Term Memory (LSTM) and VBLL LSTM to time-series sensor data collected from the High Temperature Test Facility (HTTF) at Oregon State University and train the models using data from a depressurized conduction cooldown (DCC) experiment. The deterministic LSTM and VBLL LSTM both achieve impressive predictive capabilities with ( R^2 >) 0.99 when forecasting solid and fluid temperature sensor profiles. Despite a small increase in computational cost, the VBLL LSTM is a promising direction for incorporating model uncertainty for real-time applications.

In recent years, image processing methods for gas-liquid two-phase flow, including computer vision techniques, bubble segmentation, and tracking algorithms, have seen significant development due to high efficiency and accuracy. Nevertheless, obtaining extensive, high-quality two-phase flow images continues to be a time-intensive and costly process. To address this issue, a generative AI architecture called bubbly flow generative adversarial networks (BF-GAN) is developed, designed to generate realistic and high-quality bubbly flow images through physically conditioned inputs, namely superficial gas and liquid velocities.

Initially, 105 sets of two-phase flow experiments under varying conditions are conducted to collect 278,000 bubbly flow images with physical labels of and as training data. A multi-scale loss function of GAN is then developed, incorporating mismatch loss and feature loss to further enhance the generative performance of BF-GAN. The BF-GAN’s results indicate that it has surpassed conventional GAN in generative AI indicators, establishing for the first time a quantitative benchmark in the bubbly flow domain. In terms of image correspondence, BF-GAN and the experimental images exhibit good agreement. Key physical parameters of bubbly flow images generated by the BF-GAN, including void fraction, aspect ratio, Sauter mean diameter, and interfacial area concentration, are extracted and compared with those from experimental images. This comparison validates the accuracy of BF-GAN's two-phase flow parameters with errors ranging between 2.3% and 16.6%. The comparative analysis demonstrates that the BF-GAN is capable of generating realistic and high-quality bubbly flow images for any given and within the research scope, and these images align with physical properties.

Boiling is a process in which heat from a submerged surface is removed through bubble vaporization. Consequently, the primary mode of thermal transport is governed by bubble dynamics. Therefore, one way to extract bubble data for estimating cooling performance is through the use of side-view shadowgraphs. However, linking bubble dynamics to other parameters remains challenging due to irregular bubble shapes and varied visualization setups. Even with neural networks, manual data annotation for initial model training demands significant time and effort, further complicated by instrumental differences such as varying lighting conditions. To overcome such facility limitations, this study utilizes an augmentation method to generate a synthetic bubble swarm dataset by assembling individual synthetic bubbles tailored to the experimental setup. This dataset is used to further train a Mask R-CNN for segmentation tasks to automate accurate bubble region predictions. After training, the model successfully identified bubble boundaries in high-speed camera images, extracted size distributions, and derived bubble trajectories under different heat fluxes. This demonstrates the feasibility of training models on augmented images to accurately segment bubble regions from experiments and to provide reasonable estimations of the sizes and trajectories of bubbles. This method serves as an alternative to traditional boiling heat transfer measurements, especially in setups where direct measurement of bubble statistics is limited or under conditions where conventional methods face material and instrumental constraints.

Heat transfer on boiling surfaces is a transient multi-dimensional process. Gaining an understanding of the dynamics of bubbles on such surfaces is a fundamental step toward an accurate heat partitioning models, which can be employed as best estimate models for LWR analysis. A number of traditional techniques have been developed for measuring and interpreting bubble dynamics, but they are time and computation demanding. Machine learning and computer vision have shown great potential to reduce this data interpretation burden in local feature detection such as bubble dry spot. However, a holistic segmentation of boiling features on heat transfer surfaces has not yet been demonstrated. In this study, we employed a U-NET based convolution neural network to conduct a global segmentation of a boiling surface, which allows for the classification of the dry areas, micro-sublayers, and single-phase areas from high-speed infrared images. For training the model, we obtained the ground truth partitioning from HSV phase detection images. After training, the U-NET model can predict the dry areas, micro-sublayers, and single-phase areas directly from the IR image. This successful demonstration paves the way for further research on predicting heat transfer performance based on IR images.

Thermal-hydraulic simulations of Once-Through Steam Generators (OTSGs) are crucial for operational optimization, real-time monitoring, and the development of digital twins. However, repeated and extensive simulations are computationally expensive. Model Order Reduction (MOR) techniques, such as Proper Orthogonal Decomposition (POD), offer an alternative approach to enhance computational efficiency while maintaining acceptable accuracy. Although POD is widely used for capturing dominant patterns in high-dimensional systems, its robustness within the parameter domain is limited due to its reliance on snapshots and potential inadequacy in representing highly nonlinear or complex systems. In this paper, we propose a novel framework for Parametric POD that integrates Long Short-Term Memory (LSTM) networks to enhance the accuracy and robustness of reduced-order models, and applied it to construct a Reduced-Order Model (ROM) of OTSGs. The framework leverages high-dimensional snapshots generated under varying reactor power levels, reducing them to a few dominant POD modes. Given the time-dependent nature of both the parameter (reactor power) and POD mode coefficients, LSTM is employed to approximate the mapping function between them. The resulting parametric ROM is verified for rapid estimation of primary and secondary side fluid temperatures in OTSGs using RELAP5 simulation results. The ROM achieves a maximum instantaneous fluid temperature deviation of less than 2.5 K (0.5% relative error) and reduces computation time to 1% of that required by RELAP5. This novel approach demonstrates significant potential to address computational challenges posed by numerous simulation inquiries, thereby enhancing the efficiency and applicability of OTSG modeling.