OUxSBLI is a GPU-accelerated CFD code with Python-API written in CUDA Fortran. It employs explicit high-order finite-difference schemes on a rectilinear grid (3D and 2D solvers) and a curvilinear grid.

• NVIDIA HPC SDK 24.x or 25.x — provides mpif90 and bundled MPI
• CMake 3.18 or newer
• fypp (pip install fypp) — Fortran preprocessor
• ParaView — for visualizing VTK output

• Python 3.10 or newer
• numpy
• scipy
• vtk

1. Go to a case directory (e.g. NS Taylor-Green vortex):

cd 3D_solver/NSTGV

2. Edit config.fypp to choose the physics model, convective scheme, spatial order, boundary conditions, and other compile-time options:

#:set VISC   = 'NS'      # 'Euler', 'NS', or 'LES'
#:set SCHEME = 'SLAU'    # 'KEEP', 'SLAU', or 'Hybrid'
#:set ORDER  = 6         # spatial order: 2, 4, or 6
#:set BC_X   = False     # False → periodic; True → wall/inflow BCs
#:set BC_Y   = False
#:set BC_Z   = False

See docs/configuration.md for the full reference.

3. Edit mod_globals.f90 to set grid size, domain lengths, physical parameters, and GPU thread-block sizes:

integer, parameter :: nx = 513
integer, parameter :: ny = 513
integer, parameter :: nz = 513
real(8), parameter :: Re = 1600.d0

4. Edit set.f90 if you need to change the grid geometry, initial conditions, or boundary condition routines.

5. Build with CMake:

cmake -B build && cmake --build build -j

6. Run the simulation:

cd build && mpirun -n 2 ./a.out

VTK output files (Q00000.vtr, Q00001.vtr, …) appear in the data/ directory.

Profiling: In some case directories, profile.sh runs nsys/ncu profiling:

cd build && bash ../profile.sh

1. Install the ouxsbli package:

pip install -e ".[dev]"

2. Create a Case, build, and run:

import pathlib
from ouxsbli import Case

case = Case(
    source  = "3D_solver/NSTGV",
    workdir = "/tmp/my_run",
    # physics (maps to config.fypp)
    visc    = "NS",
    scheme  = "SLAU",
    accuracy = 6,
    # grid (maps to mod_globals.f90)
    nx = 128,
    ny = 128,
    nz = 128,
)

case.build()
case.run(nranks=2)  # mpirun -n 2 ./a.out

See docs/api.md for the full API reference.

• Kinetic energy and entropy preserving (KEEP) scheme
• Simple low-dissipation AUSM (SLAU) scheme
• KEEP / SLAU hybrid scheme

• ME4-Base
• Gaitonde and Visbal's 2nd-order scheme

• Selective mixed scale model

• 3-stage TVD Runge-Kutta
• 4-stage classical Runge-Kutta

The results are consistent with Lusher's results.

• Numerical setup

@article{lusher2021assessment,
  title={Assessment of low-dissipative shock-capturing schemes for the compressible Taylor--Green vortex},
  author={Lusher, David J and Sandham, Neil D},
  journal={AIAA Journal},
  volume={59},
  number={2},
  pages={533--545},
  year={2021},
  publisher={American Institute of Aeronautics and Astronautics}
}

Wall Cp and the separation-bubble location match digitized reference data (Moro et al., 2D_solver/SBLI/ref/), verified by ouxsbli/tests/test_sbli.py.

Cf and the velocity profile match the Blasius similarity solution, verified by ouxsbli/tests/test_bl.py.

M=2 freestream with θ=8° flow deflection. Pre- and post-shock states agree with the Rankine-Hugoniot relations within 2% and 5% respectively, verified by ouxsbli/tests/test_os.py.

This repository contains the implementation used in the following publication:

Jun Hatayama, Kento Tanaka, and Toshinori Kouchi. "Nonlinear causal relationship between separation bubbles and reflected shock wave in shock wave/turbulent boundary layer interaction based on information theory." Computers & Fluids (2026): 107016.

@article{hatayama2026nonlinear,
  title={Nonlinear causal relationship between separation bubbles and reflected shock wave in shock wave/turbulent boundary layer interaction based on information theory},
  author={Hatayama, Jun and Tanaka, Kento and Kouchi, Toshinori},
  journal={Computers \& Fluids},
  pages={107016},
  year={2026},
  publisher={Elsevier}
}

The repository was made publicly available after publication to improve reproducibility. However, this version may differ slightly from the version used in the paper.

Development during 2024 and 2025 was primarily conducted by the project owner.
Starting in 2026, the project expanded its contributor base and introduced AI-assisted "vibe coding" workflows using Claude Code.

To maintain transparency, we aim to clearly distinguish which parts of the codebase and development workflow involve AI-generated content or AI-assisted modifications. In addition, as part of our effort to share practical knowledge on AI-assisted development in the HPC community, we provide Claude Code plan files under ./docs/plans.

This project is under BSD 3-Clause License