PARTICLES 2025

Accelerating SPH Simulations in LS-DYNA with Graph Network Simulator (GNS)

  • MATSUMI, Shinnosuke (Hiroshima University)
  • HATA, Toshiro (Hiroshima University)
  • HASHIMOTO, Ryota (Kyoto Unicersity)
  • SHIMAZU, Yasuhiko (KOBELCO CONSTRUCTION MACHINERY CO., LTD.)
  • Sekizuka, Ryota (KOBELCO CONSTRUCTION MACHINERY CO., LTD.)

Please login to view abstract download link

Smoothed Particle Hydrodynamics (SPH) has garnered increasing attention as a mesh-free method for simulating complex geotechnical phenomena, such as granular flows and soil–fluid interactions. Its ability to handle large deformations and discontinuities makes it particularly suitable for construction planning and earthwork simulations. However, the application of SPH in daily construction workflows remains limited due to its high computational cost. The time-consuming nature of SPH simulations hinders rapid feedback and iterative updates on construction sites. To address this issue, we propose accelerating SPH simulations by leveraging the Graph Network Simulator (GNS) [1]. Our methodology involves generating training datasets from LS-DYNA-based SPH simulations across various benchmark problems, converting them into formats compatible with GNS, and training models to predict particle interactions. A custom encoder was developed to preprocess LS-DYNA outputs for GNS input. The trained models are subsequently employed for rapid inference under similar conditions. The proposed framework was validated using standard benchmarks, including dam-break and granular column collapse scenarios. Notably, one test case follows the soil column collapse problem introduced in a recent 3D Lagrangian smoothing method study [2]. Results indicate that GNS predictions closely replicate conventional SPH behavior while substantially reducing computation time. Although our primary focus is on soil excavation, we also draw upon the work of Espinosa et al. [3], who applied SPH to high-speed metal cutting. Despite differences in material properties, their findings offer valuable insights for soil cutting simulations. Future research will extend the proposed approach to more complex interactions involving rigid bodies, such as construction machinery. We are exploring representations such as FEM-based meshes and clumped-particle DEM to incorporate rigid body effects. In addition, the framework will be expanded to predict cutting resistance and other engineering quantities essential to excavation processes.