Explainable AI for Prediction of Peak Cladding Temperature under Extended Station Blackout in a Typical Nuclear Power Plant

This study presents an interpretability analysis framework for artificial intelligence (AI) models predicting thermal-hydraulic behavior during extended station blackout (SBO) conditions in a typical nuclear reactor. The framework combines attention mechanisms, gradient-based attribution, and parameter interaction analysis to examine the model's decision-making process across accident phases. The results reveal phase-specific temporal processing patterns and evolving parameter importance, from initial power-dominated response through heat removal to stabilization. Parameter interaction matrices quantify relationships between system parameters throughout the accident progression. The gradient-based attribution analysis quantifies phase-specific parameter importance with strong attribution contrasts (±0.3), for example the strong power-pressure coupling (1.15e+00) during the initial response and primary-to-secondary heat transfer relationships (2.81e+00) during the heat removal phase. This comprehensive analysis shows alignment between AI model behavior and thermal-hydraulic principles, supporting the potential integration of such models into nuclear safety applications while maintaining transparency in their decision-making process.