Explainable critical heat flux prediction in vertical circular channels by self-attention-enhanced one-dimensional convolutional neural networks

Critical heat flux (CHF) is an important parameter in thermohydraulic systems for both nuclear and industrial applications. Given the importance of CHF, explainability in its prediction becomes necessary for those safety-critical applications like water-cooled nuclear reactors. For a predictive model to be deemed explainable, it must not only make accurate predictions but also elucidate the rationale behind specific predictions. In this paper, we propose an explainable self-attention-enhanced one-dimensional convolutional neural network (SA-1DCNN) for CHF prediction in vertical circular channels. The proposed hybrid architecture integrates convolutional layers for extracting localized patterns with self-attention mechanisms for capturing global contextual dependencies among key physical parameters. A Bayesian Optimization algorithm is used to automatically set the hyperparameters of the proposed SA-1DCNN model, resulting in an optimized, robust and generalized architecture for CHF prediction. Beyond good predictive performance, explainability and interpretability are also achieved through comprehensive analyses, including mean activation profiling, spatial attention mapping and layer-wise correlation analysis. The proposed SA-1DCNN model is validated on the experimental CHF data provided by the Working Party on Scientific Issues and Uncertainty Analysis of Reactor Systems (WPRS) Expert Group on Reactor Systems Multi-Physics (EGMUP) task force on AI and ML for Scientific Computing in Nuclear Engineering projects, supported by the OECD/NEA. The obtained results show that the proposed SA-1DCNN model not only outperforms state-of-the-art methods in predictive performance but also offers transparent and physically consistent insights into the mechanistic and thermohydraulic phenomena governing CHF. This combination of predictive power, explainability and interpretability establishes the proposed model as a promising tool for data-driven decision-making in nuclear thermohydraulic systems. • A novel explainable CHF model using a self-attention–enhanced 1DCNN was proposed. • The hybrid architectural design integrates CNN layers with self-attention mechanisms. • The model was validated on CHF experimental data from vertical circular channels. • Comprehensive analyses performed support model explainability and interpretability. • Beyond good accuracy, the model provides transparent, physically consistent insights.