Prediction of Single Channel Critical Heat Flux by Machine Learning Approaches

Abstract

Prediction of critical heat flux is crucial for the design and safety of nucleate boiling systems. Recent advancements in predicting critical heat flux center around machine learning methods, with a notable challenge being the size of the training dataset gathered from past experiments. Recently, NEA launched an international benchmarking endeavor for ML/AI based CHF modeling and prediction using the critical heat flux database provided by the US Nuclear Regulatory Commission (NRC) based on the 2006 LUT database. In this study, a large dataset covering mainly single channel CHF data with test settings and conditions similar to the NRC dataset is utilized to develop a machine learning framework for predicting critical heat flux. This machine learning framework is then applied to predict single tube CHF under various conditions, which is then compared with experimental results. The objectives include feature selection and importance analysis, model optimization, and achieving high-performance metrics. Firstly, we conducted data pre-processing including cleaning, scaling, and feature selection to ensure the high-quality datasets for training the model. Next, we developed a deep neural network (DNN) and optimized it through fine-tuning hyperparameters to train the data effectively. The overall R-squared value is outstanding for the unseen dataset, with a reasonable RMSE value. Finally, we employed the SHAP (SHapley Additive exPlanations) approach to attribute credit for the model's predictions to individual input features. The analysis suggests the potential removal of less important features to simplify the model and reduce computational complexity.