Synergistic AI for Nuclear Safety: PINNs, XAI, DTs, and LLMs in Fault Diagnosis

The operational safety and reliability of nuclear power plants (NPPs) require advanced fault diagnosis systems capable of handling rare events and complex dynamics. Traditional artificial intelligence (AI) approaches, categorized as knowledge-driven or data-driven, face limitations in safety-critical applications due to data scarcity, model opacity, and regulatory constraints. This paper emphasizes a synergistic integration of emerging AI paradigms to overcome these challenges. Physics-informed neural networks enable deep fusion of physical knowledge with data, improving generalization and robustness under limited observations. eXplainable AI enhances transparency, supports operator trust, and meets regulatory requirements by providing interpretable, actionable insights. Digital twins serve as high-fidelity virtual-physical platforms for model validation, prognostics, and scenario testing. Generative AI addresses data imbalance, while large language models unlock multimodal reasoning by processing vast unstructured operational knowledge. Together, these technologies move NPP fault diagnosis beyond pattern recognition toward systems that combine physical rigor, explainable decisions, and adaptive, context-aware intelligence, offering a pathway for safer and more reliable nuclear operations.