Bridging traditional and artificial intelligence-based models for vapour cloud hydrogen explosion hazard prediction: A critical review

Hydrogen's growing role in decarbonisation introduces new safety challenges due to its wide flammability range, low ignition energy, high diffusivity, and propensity for rapid flame acceleration. Accurate prediction of peak overpressure, positive-phase impulse, pressure–time history, and combustion regime is essential for designing hydrogen infrastructure capable of withstanding accidental deflagrations, fast deflagrations, and detonations. Traditional prediction approaches, including TNT equivalency, the TNO Multi-Energy method, Baker–Strehlow–Tang (BST), Dorofeev correlations, and computational fluid dynamics (CFD), provide valuable foundational tools but remain limited by empirical calibration, subjective parameter selection, scalability constraints, and computational expense. This review critically examines the transition from traditional empirical and numerical approaches to data-driven, hybrid, and physics-informed artificial intelligence (AI) and machine learning (ML) models for explosion prediction, with emphasis on hydrogen vapour cloud explosion applications. It evaluates how AI/ML methods can emulate, accelerate, or extend classical models while addressing data limitations, computational efficiency, uncertainty quantification, and interpretability. The review shows that current AI/ML applications for hydrogen explosions remain constrained by sparse experimental datasets, incomplete combustion regime coverage, limited validation against large-scale tests, and insufficient explainability for safety-critical decision making. It concludes by identifying key challenges in applying AI/ML to hydrogen explosion prediction and by formulating future research directions, including hydrogen-aware learning architectures, physics-informed surrogate models, unified multi-fidelity benchmark datasets, uncertainty-aware validation protocols, and integration with structural response models.