Re-examining the input-parameters and AI strategies for Critical Heat Flux prediction

This study employed three deep-learning models to predict CHF, with Transformers outperforming the other methods, thereby solidifying its leading position. The research re-examines the input parameters used in previous studies, which often relied on indirect thermohydraulic parameters, reducing prediction accuracy. By carefully selecting input parameters through mechanistic analyses and utilizing Transformer models, a minimum RMSPE of 9.85 % and NRMSE of 6.63 % was achieved using experimental data exceeding 20,000 points. This approach significantly outperformed the LUT method, which exhibited an RMSPE of 158 % and NRMSE of 21.8 %. Additionally, five traditional AI methods were tested. While most traditional methods underperformed compared to LUT, the Random Forest model achieved an RMSPE of 3.71 % and NRMSE of 4.39 %. Sensitivity to the number of input diameters was examined, showing that the overall deviation (OD) dropped from 59.75 % with a single parameter to a minimum of 2.63 % when using five parameters. It is proposed that future prediction efforts, whether using deep learning AI methods or traditional approaches, rigorously test multiple methods to identify the most accurate. Additionally, the careful selection of input parameters is crucial, as some, like inlet subcooling, can enhance accuracy more than others, such as outlet quality. • Transformer models achieve high accuracy in CHF prediction. • Random Forest outperforms some deep learning methods in CHF prediction. • Optimizing input parameters is crucial for accurate CHF prediction. • A protocol for future thermohydraulic parameter prediction is proposed.