dmipy-sim 2.0.0

dmipy-sim: GPU Monte-Carlo diffusion-MRI simulator (the forward-truth engine)

dmipy-sim

A JAX Monte-Carlo diffusion-MRI simulator — the numerical ground-truth companion to dmipy-fit. Spins random-walk through geometric substrates and accumulate phase under arbitrary free gradient waveforms G(t), with surface relaxivity and membrane permeability baked into the walk. Everything is vmap/scan JAX and runs on CPU or any CUDA-12 GPU.

The free waveform is the base representation: G(t) of shape (n_measurements, n_t, 3) is the ground truth, and PGSE/OGSE/CPMG/STE/PTE are factory constructors, not fundamental types. Magnetisation is treated as fully transverse throughout (ideal instantaneous pulses). dmipy-sim and dmipy-fit share one pulse-sequence and substrate interface — the simulator and the analytical models eat the same Waveform and the same substrate, with no conversion layer.

Quickstart — a diffusion signal in a few lines

Free (Gaussian) diffusion under a PGSE waveform reproduces the textbook Stejskal–Tanner law S/S0 = exp(-b·D) — a one-line sanity check that the install works:

import numpy as np
from dmipy_sim import simulate, pgse, set_b, FreeDiffusion

D = 2e-9   # m^2/s
for b in [0, 1e9, 2e9]:                       # b in s/m^2  (1e9 s/m^2 = 1000 s/mm^2 = 1 ms/µm^2)
    wf = set_b(pgse(delta=0.01, DELTA=0.04, G_magnitude=0.05, bvecs=[[1, 0, 0]], n_t=300), b)
    S  = float(np.asarray(simulate(n_walkers=50_000, diffusivity=D, waveform=wf,
                                    geometry=FreeDiffusion(), seed=0, require_gpu=False)).ravel()[0])
    print(f"b={b/1e9:.0f} ms/µm²   S/S0={S:.3f}   (exp(-bD)={np.exp(-b*D):.3f})")
# b=0 → 1.000   b=1 → 0.131 (exp -bD 0.135)   b=2 → 0.021 (0.018)

Swap FreeDiffusion() for a Cylinder/Sphere/PackedCylinders to see restriction, or add surface_relaxivity_t2= / permeability= to the geometry — see below. (require_gpu=False just silences the CPU-fallback warning; drop it on a GPU.)

What's here

	Entry point	Notes
Forward MC	simulate(waveform, geometry, diffusivity, ...)	walker ensemble signal Re⟨exp(iφ)⟩, with optional baked-in T2, surface relaxivity, permeability — one walk per call
Multi-echo	simulate_cpmg(n_walkers, D, cpmg_waveform, geometry, ...)	full CPMG echo train from a single walk — signal sampled at each echo (ideal 180s); (n_echoes, n_measurements)
Encodings	pgse, ogse, cpmg, ste, pte, trapezoidal_ogse	factory constructors over the free waveform; calc_b, calc_btensor, btensor_invariants
Noise	add_rician_noise, add_nc_chi_noise, estimate_sigma
Sequence I/O	dmipy_sim.sequences (incl. Pulseq .seq interop), scanner_constants	per-vendor gradient/RF/SAR deliverability limits

Geometries

Geometry	Restriction	Surface relaxivity	Permeability
FreeDiffusion	none	—	—
Box1D	1-D slab	✓	—
Sphere, Cylinder, Ellipsoid	closed wall	✓	✓
PackedCylinders, PackedSpheres	periodic ensemble	✓	✓
MyelinatedCylinder, PackedMyelinatedCylinders	multi-wall myelin geometry	✓	✓ dual-wall

Surface relaxivity & permeability

surface_relaxivity_t2=ρ (m/s) — Brownstein–Tarr: each wall collision reduces the walker weight by exp(−2ρ·d⊥/D), so the ensemble signal decays as exp(−TE·ρ·S/V) (cylinder S/V=2/R, sphere 3/R). permeability=κ (m/s) — Powles (2004) bidirectional crossing with p=min(1, 2κ·d⊥/D). Both are baked into the walk: build the geometry with the property and call simulate() (one walk per ρ/κ). The ρ weight applies on reflection only, never on transmission. Intra↔extra exchange is the Kärger/NEXI path; lipid bilayers are impermeable, so physiological exchange is at the nodes of Ranvier, not through the myelin sheath.

⚠️ Tracer self-diffusivity vs conductivity (read before benchmarking tortuosity)

The MC (and PGSE/dMRI) measure the tracer self-diffusivity D_self = MSD/4t. This is not the Fickian/effective conductivity σ_eff — they differ by the porosity: σ_eff/σ0 = (1−f)·D_self/D0. Maxwell–Garnett (1−f)/(1+f), Rayleigh, and the Hashin–Shtrikman / Wiener bounds are statements about σ_eff, not D_self. For a square array of impermeable cylinders the exact tracer value is D_self/D0 = 1/(1+f), and the MC matches it to ≤1.4%. Comparing the MC tracer diffusivity against the conductivity formula spuriously suggests a 2× error and an apparent bound violation — it is a units/quantity mismatch, not a bug.

One call, one walk

Each simulate(waveform, geometry, ...) runs a fresh spin walk and returns the ensemble signal. Surface relaxivity and permeability are substrate properties baked into that walk — set them on the geometry and call simulate():

from dmipy_sim import simulate, pgse, set_b, Cylinder

geom = Cylinder(radius=5e-6, orientation=(0, 0, 1), surface_relaxivity_t2=1e-6)
wf   = set_b(pgse(delta=0.01, DELTA=0.04, G_magnitude=0.2, bvecs=[[1, 0, 0]], n_t=300), 1e9)
sig  = simulate(n_walkers=100_000, diffusivity=2e-9, waveform=wf, geometry=geom, seed=0)

Examples

• Flagship — canonical white matter (Open in Colab) — build a histology-calibrated packed-myelinated-cylinder substrate, run the Monte-Carlo forward with surface relaxivity, and check it against the analytical model in dmipy-fit.
• Validation ladders — surface relaxivity and permeability from 1-D to 3-D vs exact analytics, and the extra-axonal tortuosity scale sweep.

Relationship to disimpy / MISST / Camino, and citation

The core engine — Brownian walk, specular reflection, phase accumulation — follows the same physics as disimpy (Kerkelä 2020), MISST, and Camino, and is validated against analytical and MISST reference signals for sphere/cylinder. Contributions here: membrane permeability (Powles), surface relaxivity (Brownstein–Tarr, interior + exterior), B-tensor encoding, packed ensembles, the free-waveform pulse-sequence interface shared with dmipy-fit, and the JAX (vmap/scan) backend.

For basic MC functionality cite disimpy/MISST; for the extensions cite dmipy-sim / the dmrai ecosystem. Kerkelä L, Nery F, Hall M, Clark C (2020), disimpy, JOSS 5(52), 2527.

Physics as the specification

dmipy-sim is written with the physical test suite as the specification — analytical solutions, eigenfunction series, Brownstein–Tarr relations, and MISST reference signals define correctness, and the code is written to pass them. Any refactor or backend change is safe as long as the suite passes.

Install

# CPU (any platform)
pip install -e ".[dev]" && pip install "jax[cpu]>=0.6.2"
# NVIDIA GPU (CUDA 12)
pip install -e ".[cuda12,dev]"

GPU note: after install, point the linker at the bundled CUDA libs in your venv's activate:

export LD_LIBRARY_PATH=$(find "$VIRTUAL_ENV/lib"/python*/site-packages/nvidia -name lib -type d | tr '\n' ':')$LD_LIBRARY_PATH

Tests

pytest -q -m "not slow"   # fast suite
pytest -q -m slow         # heavy GPU battery (first-principles validation ladders)

The tests cover: free/box/sphere/cylinder/ellipsoid diffusion vs analytical + MISST; baked-in T2; surface relaxivity (interior + exterior); B-tensor (LTE/STE/PTE); packed cylinders/spheres; permeability (all closed surfaces); SH convolution; and periodic unwrap. The slow-marked tests/validation/ battery asserts the first-principles 1D→2D→3D permeability and surface-relaxivity ladders against exact eigenvalues (see examples/validation/).

License

Dual-licensed: GNU AGPL-3.0 for open-source use, or a commercial license for proprietary/closed use. See LICENSE and LICENSING.md (commercial: [e-mail]).