flow

GPU-accelerated incompressible Navier–Stokes solver for flow in complex geometry, built around a staggered MAC grid, a signed-distance-field (SDF) description of the solid, a cut-cell Immersed Boundary Method, and a pressure-projection step with a geometric multigrid Poisson solve. The code is written in Kokkos C++ (one source runs on the CUDA, HIP/AMD, and OpenMP backends, selected at build time) and exposed to Python through nanobind (zero-copy, on core's View↔ndarray bridge); simulations are driven from Python.

The repository is also known as pnm_from_sdf (its GitLab origin) — it computes pore-network–scale flow directly from segmented SDF geometry. The pore-network extraction itself (the former peclet.flow.pnm module) now lives in its own suite project, peclet-pnm (peclet.pnm).

Modules

The original CUDA implementation has been retired (Kokkos became canonical, 2026-06); flow was validated bit-identical to the CUDA solver — to machine precision, and against Zick & Homsy sphere-array drag — before the CUDA sources were deleted (restore point: git tag pre-cuda-retirement). The shared cut-cell IBM primitives now live in src/cut_cell_ibm.hpp; the operator headers are src/mac_*.hpp + src/flow_ibm.hpp.

Capabilities

• Geometry: SDF solids (negative inside); the cut-cell IBM applies a Robust-Scaled no-slip / moving-wall condition and a matching cut-cell pressure operator (face openness from the SDF).
• Native domain boundary conditions (flow): per-face periodic / no-slip wall / Dirichlet velocity (inflow) / outflow / free-slip (symmetry plane), plus per-position inlet velocity profiles. Validated on the lid-driven cavity (Ghia et al.), the developing plane channel (Poiseuille), the backward-facing step (Armaly/Gartling), and a half channel closed by a symmetry plane (pointwise equal to the full one).
• Pressure multigrid: rediscretized geometric V-cycle, grid-independent, with MG-PCG and Chebyshev outer accelerators. Works on periodic, IBM, and non-periodic (BC) domains, including semi-coarsening for thin (quasi-2D) grids.
• Time integration: pressure projection with optional incremental pressure, explicit (Koren) or implicit-deferred-correction advection, and Picard outer iteration.

Build

# Canonical: build + install via scikit-build-core
CMAKE_PREFIX_PATH="$PWD/../extern/install/<backend>" pip install .   # -> peclet.flow
# Or a dev cmake build (nanobind found via the active interpreter, no cmakedir needed):
cmake -S . -B build -DCMAKE_PREFIX_PATH="$PWD/../extern/install/<backend>" && cmake --build build -j
# distributed flow build (opt-in MPI):
cmake -S . -B build_mpi -DCFD_BUILD_MPI=ON -DCMAKE_PREFIX_PATH="$PWD/../extern/install/<backend>" \
  && cmake --build build_mpi -j

<backend> is one of nvidia-cuda / host-openmp / lumi-hip under ../extern/install/, produced once by ../tools/bootstrap_deps.sh (a hard build dependency). Requirements: a Kokkos backend (CUDA/HIP/OpenMP — CUDA is just one option, not required), a C++20 host compiler, nanobind + scikit-build-core, and — for distributed flow — MPI. Python dependencies live in a virtual environment (.venv).

Run / verify

Simulations are scripts, not C++ mains. The scripts/verify_*_sdflow.py files are the canonical verification entry points:

source .venv/bin/activate
python scripts/verify_lid_cavity_sdflow.py     # lid-driven cavity vs Ghia, Ghia & Shin (1982)
python scripts/verify_channel_sdflow.py        # developing plane channel -> Poiseuille
python scripts/verify_bfs_sdflow.py            # backward-facing step (reattachment length)
ctest --test-dir build_mpi --output-on-failure # the multi-rank C++ test suite

Documentation

API documentation (C++ classes/kernels and Python scripts) is generated with Doxygen and published to GitHub Pages by the Documentation CI workflow. Build it locally with:

doxygen docs/Doxyfile      # output in docs/html/index.html

The architecture, conventions, and design rationale are described in CLAUDE.md and the design notes under doc/ in the repository.