Developing a memory-efficient GPGPU-parallelized contact detection algorithm for 3D engineering-scale FDEM simulations

Abstract

Despite FDEM’s superior advantages in modeling solid deformation and fracturing behaviors for engineering-scale analysis, it has been restricted to two-dimensional (2D) scenarios, with three-dimensional (3D) scenarios remaining relatively unexplored due to excessive memory constraints and computational overheads, especially in contact-related computations. This study proposes a GPGPU-parallelized contact detection algorithm to address memory constraints in 3D FDEM engineering-scale simulations. The search domain is initially divided into equally sized search cells (subCells). Through an initial element mapping procedure, subCells containing at least one solid element (realCells) are identified. RealCells are dynamically pinpointed through hashing, forming subCell-realCell lists, which allows directly allocating GPU’s memory to useful realCells instead of all subCells. Then, a second element mapping operation populates realCells with element IDs, followed by a broad search at the CUDA block level for potential contacts and a narrow search to determine actual contacts. The algorithm’s effectiveness is first verified through three benchmark tests, and then further highlighted by two engineering-scale simulations showcasing a significant decrease in GPU memory consumption with the comparable computational time. We demonstrate that the proposed algorithm reduces memory usage of the contact detection phase by 85% compared to the conventional algorithm in a 3D engineering-scale simulation involving over 8.5 million tetrahedral and 17.6 million cohesive elements.