Conference Agenda

This paper presents a summary of the results from phase 1 of the Critical Heat Flux (CHF) benchmark, organized by the OECD/NEA Task Force on Artificial Intelligence (AI) and Machine Learning (ML) for Scientific Computing in Nuclear Engineering. As the first in a series of AI/ML-focused initiatives led by the Task Force, this benchmark received 48 contributions from 31 institutions across 14 countries. Phase 1 focused on CHF regression in uniformly heated vertical tubes, utilizing a large public database of 24,579 CHF data points and a separate blind database containing 560 CHF points. Most submitted models employed either neural network architectures or gradient-boosting decision tree methods. Independent evaluations of the model predictions were conducted using standard statistical metrics applied to both databases. Model overfitting, generalization, and extrapolation performance were assessed using predictions for the blind database and various slice datasets. The results indicate that most ML models significantly outperform the 2006 Groeneveld CHF lookup table by at least a factor of 2, in part due to the explicit consideration of the heated length effects. For the considered databases, tree-based methods demonstrated superior performance, including in extrapolation scenarios to large tube diameters and high mass fluxes. While model accuracy generally improved with increasing model size (in terms of number of trainable parameters), a few promising models achieved high accuracy while maintaining a reasonable size.

Critical heat flux prediction models in water-cooled systems often rely on local equilibrium quality to avoid tracking upstream history effects. Historical critical heat flux experiments with uniformly heated vertical tubes suggest that, in sufficiently long tubes, outlet quality is an adequate parameter to substitute for the combined dependence on heated length and inlet temperature.

In this study, machine learning techniques were applied to the critical heat flux database of the United States Nuclear Regulatory Commission---the foundation for the 2006 critical heat flux look-up table of Groeneveld---to assess input parameter importance and redundancy using game-theoretic Shapley values. Supported by this analysis, three machine learning models were trained for critical heat flux prediction. Each model employed the same machine learning technique and four common input parameters: pressure, mass flow, tube diameter, and heated length. The fifth parameter varied between outlet quality, inlet temperature, and inlet subcooling.

The results confirmed previous findings that replacing parameters defining outlet conditions with those describing inlet conditions improves the statistical performance of critical heat flux prediction models. Additionally, using subcooling instead of temperature enhances predictive accuracy, particularly in cases where phase change is already occurring at the inlet.

Critical Heat Flux (CHF) plays a pivotal role in ensuring reliability and safety within boiling two-phase flow systems. Despite the development of numerous CHF prediction tools using conventional empirical correlations, machine learning, and deep learning methods, the complex mechanisms underlying CHF continue to challenge the development of a unified, accurate, and robust prediction model. The complexity is further exacerbated by varying experimental dataset developed over the decades of CHF research. In response to these challenges, the present study leverages state-of-the-art AI method, including ANN, CNN, Transformer model, and transfer learning techniques. The proposed AI-based CHF prediction model, particularly the Transformer model employing self-attention mechanisms, dynamically assigns importance to different parts of the input data. The approach significantly improves the model's capability for CHF prediction. The results indicate that the predictive performance of the Transformer-based AI model exceeds that of the Look-Up Table (LUT) method and a benchmark model from the OECD-NEA based on the database encompasses 24,579 CHF data point conducted in vertical, uniformly heated, water-cooled tubes from 59 distinct sources over the past 60 years. The five-input AI model achieved the best predictive performance: Mean P/M of 1.008, Std. P/M of 0.122, RMSPE of 12.3%, MAPE of 7.22%, NRMSE of 9.91%, and Q2 of 1.26%. Moreover, the AI-based CHF prediction model's prediction behaviors are examined and compared with the LUT method. This comparison confirms the model's resistance to overfitting. Finally, by utilizing transfer learning, the model's ability to predict CHF in tubes is extended to annulus and plate geometries.

The prediction of the Critical Heat Flux in a light-water nuclear reactor core is crucial for design, operation and safety. One of the most successful method to predict the CHF is to use the Groeneveld Look-up Table. It consists of the interpolation of more than 25000 CHF experimental data in tube geometry, recently collected in a dedicated NRC database. However, its accuracy is not completely satisfactory. To better understand and predict the CHF, the OECD/NEA Expert Group on Reactor Systems Multi-Physics (EGMUP) mandated a new task force for the development of an artificial intelligence strategy for CHF regression. In this context, the present article describes a series of machine-learning regressions applied to the NRC database to predict CHF. Trying with an increasing-complexity approach, the technics of SVR, Multilayer Perceptron, Physics-Informed Neural Network, Residual Network are applied. Almost each ML-technic gives better results than the Look-Up Table. The most performing one is the Residual Network 30x64 with a RMSPE of 11%. A sensitivity to feature selection, such as mass flow rate, pressure, diameter, length of the tube, inlet temperature, and enthalpy is performed. The length of the tube is decisive to have a good accuracy even if its physical role in the prediction is debatable.

Critical heat flux (CHF) is one of the most important parameters in the design and safety analysis of water-cooled reactors. In the past, extensive experimental studies were carried out by various researchers to understand the physical processes and to provide experimental data bases for the development of prediction models. Due to some practical reasons, such as privacy issue, experimental data obtained by one researcher cannot be made available for other researchers. This led to hundreds of prediction models with narrow valid parameter ranges.

Recently, machine learning (ML) method has attracted more and more interests in the CHF prediction. The necessary condition for a successful ML-model is a large data base for training and testing. The present study proposes a new method, called data-informed continuous machine learning (DI-CML), with the key feature to generate an artificial data base, which is almost similar to the data base of the previous researchers without knowing any original experimental data points.

This paper describes briefly the idea of the DI-CML approach, which is applied to the CHF prediction with the large CHF data base provided by the OECD-NEA benchmark working group. The results achieved so far confirm the feasibility of the DI-CML approach. At the same time, challenging issues and open tasks for the future research works are pointed out, to further develop and to improve the DI-CML approach.