A Data-Driven Approach to Critical Heat Flux: An ML-Based Method

The development of an accurate model to predict Critical Heat Flux (CHF) is essential for advancing nuclear power technology, where safety and efficiency are paramount. In this work, we present a Machine Learning (ML)-based model constructed on the latest NEA benchmark dataset for CHF prediction. Our methodology incorporates comprehensive analyses of feature selection and engineering to enhance model learning capacity, alongside a novel data augmentation strategy combining Gaussian Noise Injection (GNI) and Mixup techniques with physical constraints to ensure generation of realistic synthetic data. Results demonstrate that our 7-layer feedforward neural network with batch normalization and optimized dropout layers significantly outperforms traditional empirical models and Look-Up-Tables (LUTs) in regression tasks. The augmented model achieves reduced error metrics across all performance indicators, with Mean Absolute Percentage Error decreasing from 8.51% to 7.87% and Normalized Root Mean Square Error improving from 10.18% to 8.49%. Notably, our approach better captures known physical dependencies—such as the monotonic relationship between CHF, pressure, and mass flux particularly in underrepresented parameter regimes. This data-driven architecture effectively leverages hidden relationships in the CHF dataset while maintaining physical consistency, offering a more adaptable and precise tool for predicting CHF in nuclear reactor cooling system design and operation.