MRSimTracks

Generate CFD-derived particle trajectories for MR flow simulation.

MRSimTracks performs Lagrangian particle tracking in time-resolved (pulsatile) CFD meshes. It operates on mesh velocity fields, not MR image data. It seeds particles in a tetrahedral flow domain, advects them through a time-periodic velocity field (RK4), and recycles out-of-bounds particles back to the inflow boundaries with optional backflow-aware reseeding.

Install

MRSimTracks is published on PyPI:

uv add "mrsimtracks==0.2.0"

or with pip:

python -m pip install "mrsimtracks==0.2.0"

To install the latest source from GitHub instead:

uv add "mrsimtracks @ git+https://github.com/mcgrathcm/MRSimTracks.git"

For development from a clone:

uv sync          # installs the package (editable) and dependencies

The distribution name and Python import package are both mrsimtracks.

Quick start

import numpy as np

import mrsimtracks as mt
from mrsimtracks.seeding import seed_mesh

# 1. Load a time-resolved flow field (.vtu single-file series or .pvd collection)
flow = mt.load_flow("case.pvd", active_key="Velocity")

# 2. (optional) Backflow-aware inflow reseeder from labeled cap surfaces
reseeder = mt.BoundaryReseeder(["Inlet.vtp", "Outlet.vtp"], flow, dt=0.002)

# 3. (optional) Near-wall no-penetration projection
wall_slip = mt.WallSlip(flow, caps=["Inlet.vtp", "Outlet.vtp"])

# 4. Seed and track
seeds = seed_mesh(flow.active_mesh, 200_000, rng=np.random.default_rng(0))
result = mt.track(
    flow,
    seeds=seeds,
    dt=0.002,
    reseeder=reseeder,
    wall_slip=wall_slip,
)

# 5. Use / save
result.positions      # (n_steps, n_particles, 3)
result.reset          # (n_steps, n_particles) reseed flags
result.times          # (n_steps,)
result.save("tracks.h5")

For large single-process runs, write each timestep directly to HDF5 instead of keeping all positions in RAM:

result, metrics = mt.track(
    flow,
    seeds=seeds,
    dt=0.002,
    reseeder=reseeder,
    output_path="tracks.h5",
    return_metrics=True,
)

result.is_file_backed  # True until result.positions or result.reset is loaded
metrics["particle_steps_per_s"]

Run example.py for a complete version using the reduced fixture committed in tests/data/. The full-cycle example flow file is tracked with Git LFS; see example/README.md.

Large runs (multiple processes)

Each worker reloads the field, so memory scales with n_workers:

result = mt.track_parallel(
    "case.pvd", seeds=seeds, dt=0.002,
    caps=["Inlet.vtp", "Outlet.vtp"], active_key="Velocity",
    n_workers=3, subsamp=1, wall_slip=True,
)

Key functions

Reseeding notes

• The reseeder weights every cap face by max(-v·n, 0)·area at the nearest flow frame, so particles re-enter only through currently-inflow faces — handling backflow and partial inflow/outflow on a single cap.
• BoundaryReseeder(..., dt=dt) spreads new particles over a thin inflow volume (depth ~v_n·dt) so successive reseeds overlap, keeping spatial density smooth (important for MR-style uniform-density use). Omit dt for plane seeding.
• flux_waveform() returns per-cap net flux over the cycle — a conservation / validation diagnostic (Σ caps ≈ 0 for a well-resolved incompressible field).

Wall-slip notes

WallSlip removes only the into-wall velocity component for particles within a thin wall band. Use it when interpolated near-wall velocities deposit particles against no-slip walls. Pass the same cap surfaces used for reseeding so open boundaries are excluded from the wall set:

wall_slip = mt.WallSlip(flow, caps=["Inlet.vtp", "Outlet.vtp"], band_frac=0.02)
result = mt.track(
    flow,
    seeds=seeds,
    dt=0.002,
    reseeder=reseeder,
    wall_slip=wall_slip,
)

For track_parallel, set wall_slip=True; each worker builds its own projection from the worker-local flow and caps.