Development of a data-driven turbulence model for coarse-grid CFD simulation of hydrogen dispersion in large spaces

Hydrogen safety in reactor containments is a critical issue, because hydrogen released during accidents may accumulate in large spaces and create the risk of combustion or explosion. Fine-grid CFD (FG-CFD) simulation can accurately predict hydrogen dispersion, but the high computational cost limits its application to transient processes in large spaces. In this study, a data-driven turbulence model was developed and used to construct an acceleration method, using coarse-grid CFD (CG-CFD) simulations. A high-fidelity FG-CFD database (Gr = 3.02 × 10$^{10}$–1.90 × 10$^{15}$) was built, and a mapping guideline was proposed to transfer fine-grid data to coarse grids. Based on this database, a machine learning Reynolds stress (MLRS) model was developed using deep neural network (DNN) and integrated into OpenFOAM to replace traditional turbulence models. Validation results show that the proposed method significantly reduces computational cost while maintaining high accuracy in predicting hydrogen concentration fields and exhibits good generalization capability for unseen cases within the considered range of Gr numbers. This work provides a feasible approach for efficient and reliable analysis of hydrogen dispersion in large spaces to support containment safety evaluation.