GPU optimized integration of immersed boundary method and overset mesh framework for moving boundary problems

This paper presents a GPU-accelerated overset immersed boundary framework for the high-fidelity simulation of incompressible flows involving multiple moving bodies and complex geometries. The proposed methodology combines the simplicity and flexibility of Cartesian immersed boundary methods with the local refinement capability of overset grids, enabling accurate resolution of flow features around dynamically moving bodies while maintaining computational efficiency. To support overset mesh movement due to arbitrary body motion, a lightweight sliding-window reallocation strategy is developed in conjunction with minimal interface-storage requirements and flux-preserving interpolation between coarse and refined meshes, thereby avoiding the expensive hole-cutting and donor-search procedures commonly associated with conventional overset approaches.

A hybrid Array-of-Structures (AOS) data organization is employed to efficiently manage the hierarchical overset-grid data, while CUDA Unified Memory simplifies the handling of dynamically evolving block connectivity and avoids the complex device-memory management typically required for hierarchical overset data structures. The framework is accelerated using OpenACC and extended to multi-GPU architectures through a dual-level OpenMP–OpenACC parallelization strategy, in which individual overset blocks are assigned to dedicated GPUs.

The solver is verified and validated through a broad range of benchmark problems, including flow past cylinders and spheres, sedimentation of smooth and rough particles, and flutter-induced motion of rigid plates. The framework is further demonstrated on challenging multi-body configurations involving multiple rotating drone propellers and independently flapping robotic butterflies following prescribed sinusoidal trajectories. The overset refinement strategy enables DNS-level resolution in the vicinity of moving bodies while significantly reducing the overall computational cost relative to uniformly refined meshes.