Deployment of Traditional and Hybrid Machine Learning for Critical Heat Flux Prediction in the CTF Thermal Hydraulics Code

TL;DR

The paper demonstrates native deployment of both a pure data-driven neural network and two physics-informed hybrids (base correlations plus ML residuals) for critical heat flux prediction inside the CTF subchannel code. By training on the NRC CHF database and validating against the Bennett DO experiments, the study shows that hybrid models reduce error and mitigate overprediction relative to traditional correlations, while the pure ML model remains highly competitive. The results indicate that integrating ML within a physics-based framework can enhance predictive accuracy and generalization in nuclear thermal hydraulics, with minimal runtime impact. The work also outlines a practical workflow for in-code ML deployment and highlights directions for uncertainty quantification and transferability to other geometries.

Abstract

Critical heat flux (CHF) marks the transition from nucleate to film boiling, where heat transfer to the working fluid can rapidly deteriorate. Accurate CHF prediction is essential for efficiency, safety, and preventing equipment damage, particularly in nuclear reactors. Although widely used, empirical correlations frequently exhibit discrepancies in comparison with experimental data, limiting their reliability in diverse operational conditions. Traditional machine learning (ML) approaches have demonstrated the potential for CHF prediction but have often suffered from limited interpretability, data scarcity, and insufficient knowledge of physical principles. Hybrid model approaches, which combine data-driven ML with physics-based models, mitigate these concerns by incorporating prior knowledge of the domain. This study integrated a purely data-driven ML model and two hybrid models (using the Biasi and Bowring CHF correlations) within the CTF subchannel code via a custom Fortran framework. Performance was evaluated using two validation cases: a subset of the Nuclear Regulatory Commission CHF database and the Bennett dryout experiments. In both cases, the hybrid models exhibited significantly lower error metrics in comparison with conventional empirical correlations. The pure ML model remained competitive with the hybrid models. Trend analysis of error parity indicates that ML-based models reduce the tendency for CHF overprediction, improving overall accuracy. These results demonstrate that ML-based CHF models can be effectively integrated into subchannel codes and can potentially increase performance in comparison with conventional methods.

Figures (8)

• Figure 1: Two-component PCA paired with convex hull analysis visualizing the evaluation data in an interpolating regime. The NRC test dataset's distribution is similar to that of the training data, with coverage extending to the extrema. The Bennett test series are well enclosed by the boundary and are located in a region of dense training data.
• Figure 2: Workflow for the hybrid model methodology in the training configuration. Instead of directly predicting CHF values, the ML model component predicts the difference (residual) between the base model's CHF output and the experimental CHF value. After training, in the deployment configuration, actual residual values are unknown, so it is entirely up to ML model to adjust the base model's output only with knowledge of the corresponding input vector.
• Figure 3: Comparisons between the baseline CHF models and the pure ML predictions using the test partition of the NRC dataset. The parity plot is windowed at 6,000 kW□m for ease of interpretation, and the KDE plot is windowed at $\pm 70\%$.
• Figure 4: Comparisons between the baseline Biasi CHF model and its hybrid counterpart using the test partition of the NRC dataset. The parity plot is windowed at 6,000 kW□m for ease of interpretation, and the KDE plot is windowed at $\pm 70\%$.
• Figure 5: Comparisons between the baseline Bowring CHF model and its hybrid counterpart using the test partition of the NRC dataset. The parity plot is windowed at 6,000 kW□m for ease of interpretation, and the KDE plot is windowed at $\pm 70\%$.