37x Speedup in Lattice Boltzmann Cylinder Flow

· ai design · Source ↗

TLDR

  • Paper shows 9x grid coarsening in D2Q9 LBM cylinder flow preserves Strouhal number within 2.5% while achieving 37x wall time speedup at Re=100.

Key Takeaways

  • Vortex shedding frequency (St) is resolution-robust: coarsest grid (35,511 cells, 81s) matches DNS (320,000 cells, 2958s) within 2.5% error.
  • Mean drag coefficient degrades moderately (1.8% error at 3x coarse) but stays within literature range, meaning force amplitude is the real coarsening cost.
  • Implication for sub-grid model design: target force amplitude recovery, not frequency recovery, since frequency is already captured by large-scale geometry and Re.
  • Known limitation: absolute St is 18.7% below literature due to 10% blockage ratio; results are 2D only, Re=100 only.
  • Part of the KPBM framework; goal is 37x speedup with DNS-level accuracy via nodal stability checks at high-shear interfaces.

Hacker News Comment Review

  • Thin discussion; one commenter flagged interest in whether agents assisted with the math and noted the sampling-at-privileged-moments insight as agent-friendly to test.

Original | Discuss on HN