Abstract
In subcooled boiling flow under low-pressure conditions, the bubble characteristic diameter significantly influences the surface heat transfer coefficient. Traditional mechanistic models and empirical correlations often exhibit large errors across diverse experimental conditions. This paper proposes a data-driven model using artificial neural networks (ANN) to predict the bubble maximum diameter ($D_{max}$) and investigates the impact of various experimental conditions. Established and validated using six experimental databases, the ANN model achieves a relative error of approximately 14%. Uncertainty analysis identifies pressure as a highly sensitive parameter, particularly for bubbles larger than 0.5 mm. Based on these findings, a new correlation for coefficients $C$ and $\phi$ is proposed, which improves accuracy for large bubble datasets and maintains a low relative error in independent validations.
Introduction
In nuclear reactors, specifically Pressurized Water Reactors (PWRs), subcooled boiling occurs when bubbles form at the heated surface while the bulk liquid remains below its boiling point. Predicting the bubble maximum diameter ($D_{max}$) is essential for determining the surface heat transfer coefficient and preventing the "boiling crisis"—a state where heat transfer efficiency drops sharply, risking fuel cladding damage.
Existing mechanistic models, which rely on physical heat balances, often struggle with accuracy, showing relative errors as high as 40% across different experimental setups. This research bridges that gap by employing a data-driven approach to enhance prediction reliability.
Methodology
The study utilizes a two-layer feed-forward Artificial Neural Network (ANN) optimized with a Back-Propagation (BP) algorithm and Bayesian regularization (trainbr). This optimization method is robust for smaller, noisy datasets as it prevents overfitting by minimizing a combination of squared errors and weights.
Input Parameters and Architecture
The model incorporates two hidden layers with eight neurons each. Input parameters include both raw experimental values (pressure, heat flux, mass flux) and non-dimensional numbers (Reynolds, Prandtl, Peclet, Eotvos, and Morton numbers) to capture complex fluid dynamics.
Hybrid Approach
Beyond the ANN, the researchers used sensitivity results from the network to refine traditional physics-based models. They proposed a new correlation for the coefficients $C$ and $\phi$, which represent pressure and fluid dynamic effects, respectively, replacing them with a unified $C\phi$ term driven by Peclet ($Pe$) and Eotvos ($Eo$) numbers.
Datasets and Validation
The model was trained and validated using six independent experimental databases (Prodanovic et al., Situ et al., Brooks et al., Ahmadi et al., Okawa et al., and Xu et al.), comprising 366 data points. These datasets primarily cover low-pressure and low-velocity conditions. An additional independent dataset (Kaiho et al., 2017) served as a final verification for the new mechanistic correlation.
Key Results and Insights
- Superior Accuracy: The ANN model achieved a relative error of 13.54%, a significant improvement over traditional models like the Ünal model, which can exceed 100% error in some cases.
- Pressure Sensitivity: Uncertainty analysis revealed that pressure is the most sensitive parameter for predicting bubble size, especially for bubbles exceeding 0.5 mm. Pressure fluctuations alter vapor density, directly impacting bubble volume.
- Improved Physics: The new mechanistic correlation derived from ANN insights reduced the error for large bubbles (>1 mm) from 51.39% to 33.19%.
- Robustness: When tested against the independent Kaiho dataset, the improved model maintained a low relative error of 14.43%, demonstrating its generalizability.
Conclusion
While the ANN model provides high accuracy, its "black box" nature limits its application to the range of experimental conditions it was trained on (low pressure and velocity). Future research will focus on expanding datasets to include higher pressure and mass velocity conditions and further refining hybrid, physics-informed machine learning models to enhance transparency and safety in nuclear applications.
Summary updated