Critical heat flux (CHF) prediction faces a persistent trade-off between predictive accuracy and physical interpretability in nuclear thermal-hydraulic engineering. Traditional empirical correlations suffer from narrow applicable ranges, while data-driven AI models generally function as black boxes with limited mechanistic transparency.

This study develops an interpretable CHF prediction model via PySR symbolic regression, trained on a wide-range OECD/NEA experimental database containing 24,579 measurements. The optimal model yields a mean absolute percentage error of 11.56% — on par with the classical lookup table and notably superior to conventional empirical correlations. It also maintains robust performance under input noise and out-of-range extrapolation, with its parameter dependencies consistent with well-recognized physical mechanisms.

The proposed approach offers a practical, transparent alternative that bridges oversimplified empirical formulas and opaque machine learning models.