Abstract: Under the influence of ground motion loads, saturated soil is susceptible to liquefaction, leading to secondary disasters such as ground settlement, foundation failure, and sand ejection with water seepage. When performing dynamic response analysis of saturated media using the traditional central processing unit (CPU)-based serial numerical calculation method, the large number of coupled model coefficient matrices and the strong coupling between equations result in issues such as low computational efficiency and precision. To address the numerical computing bottleneck associated with ultra-large-scale complex fluid-structure coupling models under the traditional numerical framework, this study proposes a computational framework based on CPU and graphics processing unit (GPU) heterogeneous parallelism. By optimizing the parallel computing process for large-scale data, minimizing data transfer between the CPU and GPU, and employs a data packaging strategy to reduce unnecessary time consumption. By directly constructing the global matrix in compressed sparse row (CSR) compression format through a preprocessing non-assembly method, it avoids the significant memory consumption typically incurred by traditional assembly methods. This enables the program to compute ultra-large-scale finite element models with reduced memory resources, thereby lowering both memory and time costs. Leveraging an intrinsic CUDA function, a novel set of iterative solution schemes for equations was developed, and a parallel iterative solver tailored for multi-field coupling problems was constructed. For complex fluid-structure interaction dynamic response problems, the proposed parallel solver demonstrates an order of magnitude improvement over traditional serial computation methods. The computational capacity threshold surpasses tens of millions of degrees of freedom. Compared to ABAQUS software, the computational efficiency has increased by more than 15 times.

Key words: finite element, GPU parallel computing, fluid-structure coupling, dynamic response