Integrating artificial intelligence into nuclear reactor safety: ML-augmented CFD, AI-enabled PSA, and cybersecurity

This review provides a structured synthesis of how artificial intelligence (AI) is reshaping safety-critical analysis in nuclear power plants through three complementary technical domains and cross-cutting enablers. First, we examine machine learning (ML) augmentations to computational fluid dynamics (CFD) and reduced-order thermal–hydraulic models, focusing on boiling and critical heat flux (CHF) closures, interfacial transport, transient emulation, and containment-scale mixing. Particular attention is given to physics-constrained surrogates, verification, validation, and uncertainty quantification (VVUQ), and best-estimate plus uncertainty (BEPU) practices that enable credible acceleration of high-fidelity analyses. Second, we review AI integration within probabilistic safety assessment (PSA), covering Bayesian and hybrid graphical models, dynamic event and fault trees, human reliability analysis enhanced by learning-based support, and digital-twin frameworks that embed online updating of risk metrics. Third, we examine cybersecurity in digital instrumentation and control (I&C) systems, including threat pathways, AI-enabled detection and response, adversarial model robustness, and the coupling of cyber–physical consequences with PSA. These technical strands are complemented by recent developments in uncertainty quantification and resilient AI, together with a dedicated assessment of multi-unit and multi-hazard PSA covering inter-unit and inter-module dependencies, shared structures, systems, components, and recovery resources, multi-unit accident sequences, and single, combined, and cascading external hazards. Across all domains, we consolidate comparative evidence, highlight regulatory and explainability requirements, and identify open challenges linked to data scarcity, plant heterogeneity across small modular reactor (SMR), pressurized water reactor (PWR), and boiling water reactor (BWR) designs, and physics–data integration. The outcome is a consolidated map of methods and datasets, together with a lifecycle governance framework for AI-enabled PSA that links intended use, data and model qualification, PSA integration, configuration control, operational monitoring, requalification, and human oversight to auditable decision gates.