Answers are grounded in this paper and cite their sources.
BackgroundThe post-dryout (PDO) heat transfer coefficient plays a critical role in thermal safety assessment in nuclear systems. However, accurate modeling of PDO heat transfer remains challenging due to experimental data scarcity and pronounced distribution shifts across datasets from different sources.PurposeThis study aims to develop a robust data-driven prediction framework that can improve the generalization and extrapolation performance of PDO heat transfer coefficient models under limited and distribution-shifted data conditions.MethodsA two-stage data augmentation and prediction approach based on Conditional Tabular Generative Adversarial Network (CTGAN) was proposed. In the first stage, CTGAN was trained to learn the joint distribution of dimensionless input features that are highly correlated with the Nusselt number (Num), and synthetic feature samples were generated accordingly. In the second stage, a supervised regression model based on a backpropagation neural network (BPNN) was employed to predict the corresponding target values, thereby decoupling input feature generation from input-output mapping. The generated synthetic data were then combined with real experimental samples to train the final predictive model. Finally, the proposed method was evaluated on four independent post-dryout heat transfer datasets and compared with other data generation techniques.ResultsThe experimental results show that the proposed two-stage approach significantly reduces prediction errors compared with models trained using only real data. CTGAN achieves the lowest mean absolute percentage error (MAPE) on the Becker and SWE datasets and outperforms other generation models, including tabular variational autoencoder (TVAE), tabular generative adversarial network (TGAN), and Conditional Tabular Generative Adversarial Network Plus (CTABGAN+). Compared with traditional one-stage data generation schemes, the proposed method exhibits superior extrapolation capability and robustness under distribution shift conditions.ConclusionsThe proposed CTGAN-based two-stage data generation framework effectively enhances the predictive accuracy and generalization ability of PDO heat transfer coefficient models, hence providinga reliable and extensible solution for data-scarce and distribution-shifted thermal-hydraulic modeling problems in nuclear engineering applications.
No takes yet. Share an insight, caveat, or question.
YUAN et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: