A comprehensive GPU-accelerated lattice Boltzmann framework for high-Reynolds-number flow in complex rod bundles

Rod bundles are vital components in nuclear reactors, where flow behaviour directly affects operational safety. Accurate and efficient simulation of such flows is therefore essential. However, traditional computational fluid dynamics methods often require structural approximations and cannot directly handle complex features such as contact points in self-supporting helical cruciform (HC) rod bundles, due to meshing challenges and high computational cost. To overcome these challenges, this study presents hreoLB, a high-fidelity Lattice Boltzmann framework for simulating high-Reynolds-number flows in rod bundles. It features an innovative geometry preprocessing method for automatically handling intricate geometries and large-scale grids, and a high-order LB model optimised for GPUs to ensure stable and accurate turbulent flow simulations. With support for heterogeneous computing interfaces and multi-level communication overlapping, hreoLB offers excellent scalability and portability. The framework is validated using canonical benchmarks including the Taylor–Green vortex and flow past a cylinder, as well as realistic rod bundle with spacer grids. hreoLB is applied to a [Formula: see text] HC rod bundle, first reproducing the traditional gap-based setup, followed by the first simulation of a physically contacted configuration. The results reveal a rotationally symmetric U-shaped vortex downstream of the contact points, as well as a flattening effect of flow sweeping mixing rate in the gapped model. Scalability tests show up to 70% efficiency on 10,000 CPUs and near-perfect GPU scaling, completing a half-billion-node case in 5.5 hours with 16 GPUs.