Exascale-ready HemeLB for large-eddy simulation of full-human arterial blood flow
Abstract
Direct numerical simulation (DNS) of blood flow through the whole human arterial tree is limited by the number of lattice sites that must be advanced on a high-performance computer. We address this limit by combining the lattice Boltzmann solver HemeLB with Smagorinsky large-eddy simulation (LES). HemeLB runs with the same scaling behaviour on AMD, Intel and NVIDIA graphics processing units (GPUs) and sustains about 10 trillion lattice-site updates per second (TLUPS) on both the Frontier and Aurora exascale supercomputers. In stenotic aortic flow, LES on lattices five and ten times coarser than a 20~μm DNS reproduces the wall pressure and the cross-sectional pressure and velocity, and its plane-averaged pressure differs from that of the DNS by at most 0.75~mmHg, below 1\% of the local pressure. At 100~μm, LES reduces the estimated cost of a full-human simulation by a factor of about 625 relative to DNS. We use this resolution to compute blood flow from the head to the feet in a full-human arterial model with 295 million fluid sites, one inlet at the aortic root and 432 Windkessel outlets, at about 15--18 node-hours per cardiac cycle on two Frontier nodes. A single simulation resolves every region simultaneously, from the cerebral arteries to the arteries of the feet. The pressure pulse grows as it travels from the aorta to the feet, the shear stress concentrates in the branches rather than in the aorta, the pulse reaches the outlets later with increasing distance from the heart, and the flow in the legs and feet is triphasic, as in Doppler measurements of the human leg. These results bring patient-specific whole-body lattice Boltzmann simulations within reach of current supercomputers and pave the way for ensemble simulations towards a standard human model.
Related
- Towards exascale fully relativistic pseudopotential density functional theory calculations enabled by mixed-precision computation and compressed-communication using residual based subspace iteration
- Accelerating phase-field simulations on exascale computing systems for faster-than-real-time precipitate aging predictions
- Large-scale NMR simulation on a trapped-ion quantum computer
- Towards Heterogeneous Exascale CFD with a Single Fortran Code Base: OpenMP Target Acceleration of the High-Order Unstructured Solver UCNS3D
- Large integration time-step in molecular dynamics simulation artificially enhances the strength of hydrophobic interaction
BibTeX
@misc{xue2026exascaleready,
title = {{Exascale-ready HemeLB for large-eddy simulation of full-human arterial blood flow}},
author = {Xiao Xue and Ioannis Zacharoudiou and Peter V. Coveney},
howpublished = {arXiv},
year = {2026},
month = nov,
doi = {10.48550/arxiv.2611.07840},
eprint = {2611.07840},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2611.07840},
}
RIS
TY - UNPB
TI - Exascale-ready HemeLB for large-eddy simulation of full-human arterial blood flow
AU - Xiao Xue
AU - Ioannis Zacharoudiou
AU - Peter V. Coveney
PY - 2026
DA - 2026/11/05
PB - arXiv
DO - 10.48550/arxiv.2611.07840
UR - https://arxiv.org/abs/2611.07840
AB - Direct numerical simulation (DNS) of blood flow through the whole human arterial tree is limited by the number of lattice sites that must be advanced on a high-performance computer. We address this limit by combining the lattice Boltzmann solver HemeLB with Smagorinsky large-eddy simulation (LES). HemeLB runs with the same scaling behaviour on AMD, Intel and NVIDIA graphics processing units (GPUs) and sustains about 10 trillion lattice-site updates per second (TLUPS) on both the Frontier and Aurora exascale supercomputers. In stenotic aortic flow, LES on lattices five and ten times coarser than a 20~μm DNS reproduces the wall pressure and the cross-sectional pressure and velocity, and its plane-averaged pressure differs from that of the DNS by at most 0.75~mmHg, below 1\% of the local pressure. At 100~μm, LES reduces the estimated cost of a full-human simulation by a factor of about 625 relative to DNS. We use this resolution to compute blood flow from the head to the feet in a full-human arterial model with 295 million fluid sites, one inlet at the aortic root and 432 Windkessel outlets, at about 15--18 node-hours per cardiac cycle on two Frontier nodes. A single simulation resolves every region simultaneously, from the cerebral arteries to the arteries of the feet. The pressure pulse grows as it travels from the aorta to the feet, the shear stress concentrates in the branches rather than in the aorta, the pulse reaches the outlets later with increasing distance from the heart, and the flow in the legs and feet is triphasic, as in Doppler measurements of the human leg. These results bring patient-specific whole-body lattice Boltzmann simulations within reach of current supercomputers and pave the way for ensemble simulations towards a standard human model.
ER -