High-fidelity void-fraction signals constitute essential data for modeling two-phase flows. However, the scarcity of such data constrains the development and validation of high-accuracy models, thereby impeding the design and optimization of complex industrial systems, including nuclear reactors and other energy facilities. To address this challenge, this study proposes a novel database enhancement framework, termed VoidGAN, based on conditional generative adversarial networks (GANs). The proposed model integrates Transformer modules with multi-scale convolutional Inception blocks, enabling it to capture both long-term temporal dependencies and local, irregular fluctuations. In addition, a physics-metrics-guided Bayesian hyperparameter optimization strategy is introduced to enhance the physical fidelity of the generated signals. A comprehensive multi-step validation framework is further established to rigorously assess the reliability of the generated data, encompassing direct comparisons with testing datasets and benchmarking against established mechanistic models, including the two-group drift-flux model and the two-phase flow-induced vibration (TP-FIV) excited force model. The results demonstrate that VoidGAN achieves the best overall performance among state-of-the-art time-series generative models, attaining a recall exceeding 99.8%, achieving the lowest nearest-neighbor distance (0.069), and maintaining inference times at the millisecond scale. These results confirm that both time-averaged and temporal characteristics, as well as their intricate relationships across diverse flow regimes, are accurately captured. This work provides a new perspective for mitigating data scarcity issues in two-phase flow modeling and paves the way for more efficient design and optimization of industrial systems.
扫码关注我们
求助内容：
应助结果提醒方式：
京公网安备 11010802042870号