Hugging Face's logo
Activity Feed

AI & ML interests

None defined yet.

Recent Activity
rfick  updated a dataset 3 days ago
SubstrateCommons/test-fixtures
rfick  updated a dataset 4 days ago
SubstrateCommons/canonical-pores
rfick  updated a dataset 4 days ago
SubstrateCommons/winther-g6-axons
View all activity

Organization Card
Community About org cards

SubstrateCommons

A library of replayable Monte-Carlo substrates for MRI microstructure. Each dataset here is a family of replay packs: one substrate, walked once at a certified precision, stored so that any acquisition can be replayed on it later, on your machine, in seconds. A pack carries the walkers' paths in a band-limited form, their compartment occupancy, their wall contact, and the substrate's internal field, plus the specification of the substrate it came from and a certificate stating the Monte-Carlo floor and the codec error per tier. The tissue (T2, T1, surface relaxivity, susceptibility), the scanner's field and the pose are knobs applied at replay, never baked into the walk.

Every family credits the people whose substrate it is: the source, its licence as stated by its host, and the published measurement or simulation the family reproduces, which its card shows beside our number.

Use a pack

# pip install dmipy-sim
from dmipy_sim.replay import ReplayPack
from dmipy_sim import sequences
pack = ReplayPack.load('hf://SubstrateCommons/imperial-rocks/packs/f42a.rpk')
seq = sequences.pgse([[1, 0, 0]], 0.04, 0.08, bvalues=[1e9], TE=0.15)
S = pack.replay(seq)                                              # the bare diffusion signal
S = pack.replay(seq, tissue=pack.nominal, scanner=pack.nominal_field_T)   # the paper's replay

A replay refuses what the pack cannot serve (a waveform beyond its temporal band, a tier it does not carry) rather than returning a plausible number. The format is specified in replay-pack-spec; the engine and the producers are dmipy-sim.

Families

family what it is source reference it reproduces packs licence
canonical-pores Single cylinders, spheres and planes from 0.1 to 20 µm, one walk per size at a certified floor: the closed-form references every replay is checked against, and the compartments dmipy-fit's replay models read. dmipy-sim's analytic geometries; the fit-side models are Fick et al., dmipy (Front. Neuroinform. 2019). analytic (the restricted-diffusion series of each shape) 600 (25.7 GB) CC-BY-4.0 (ours)
disco-replay The DiSCo 1 mm³ strand phantom (12,196 strands) walked in voxel blocks and stored as a columnar image: one voxel, one waveform or the whole volume, read by reference. Rafael-Patino et al., Data in Brief 38:107429 (2021), doi:10.1016/j.dib.2021.107429. A: the released MC/DC DWI volume (4 shells, 364 measurements) on the same strands 1 CC-BY-4.0
imperial-rocks Micro-CT sand packs walked on their own voxels through the label-volume producer: the surface-relaxation reference outside the brain, replayed as a CPMG decay. Dong & Blunt 2009 images (Imperial College, Figshare); Talabi 2008 thesis for the random-walk and measured CPMG. A: Talabi's random walk on the exact released images and measured CPMG on the same sand (relaxivity fitted by him) 2 CC-BY-4.0 images; thesis CC-BY-ND (numbers cited)
ling-sand-packs Quartz and garnet sand packs imaged by micro-CT after the CPMG was measured on the same resin-set packs: the same-object rock reference. Ling et al., Transp. Porous Media (2022), doi:10.1007/s11242-022-01764-w; figshare 10.6084/m9.figshare.17161730. B: measured CPMG on the same packs, released as raw trains; the paper's own random walk is a figure, not a number; 3 of 5 packs are two-mineral mixtures the contact channel cannot yet serve (blocked) 5 CC-BY-4.0
parity-fixtures Cross-engine fixtures: single undulating axons released by MC/DC with the raw DWI their engine produced, and Disimpy's cylinder against the MISST matrix-formalism signal. Rafael-Patino et al., Front. Neuroinform. 2020 (Robust-Monte-Carlo-Simulations, LGPL-2.1); Kerkelä et al., JOSS 2020 (Disimpy, MIT). A: their released signals at their scheme, compared per measurement within the two Monte-Carlo floors 3 LGPL-2.1 / MIT sources; packs CC-BY-4.0
winther-g6-axons 29 real axons segmented from X-ray nano-holotomography of a vervet brain, walked with their myelin sheath as a susceptibility source: positions, occupancy, wall contact and the field's path integral in one pack each. Winther et al., Sci. Rep. 14:29636 (2024), doi:10.1038/s41598-024-79043-5; morphology from resources.drcmr.dk. A: per-axon Monte-Carlo signals and B-fields released with the paper; ex vivo PGSE on the same tissue block 29 (2.7 GB) none stated on the host (cite the paper); packs CC-BY-4.0

Reference grades follow the survey in dmipy-sim#459: A a published number on the same released geometry, B a published figure, analytic a closed form. Each family's card carries the numbers side by side and a reproduction.json they are rendered from.

How a family is made

One producer reads the released substrate (a mesh, a label volume, a strand list, a sphere set, an analytic shape) into a substrate specification; one walk at a floor set by a measured pilot; one pack per substrate with its certificate; a gate that reads the recorded reference numbers and refuses to publish a pack that does not reproduce them; a card generated from the records. The protocol is dmipy-sim#482, the card dmipy-sim#480.

Contribute, request, or object

Open an issue on dmipy-sim: to propose a released substrate for a family (say where it is, its licence, and what was measured on it), to request an acquisition or a tier on an existing family, or, if you hold rights in a source and want a family taken down, to say so; the card is deleted on request.

Cite

The Commons and the replay method are described in the replayable Monte-Carlo paper (Fick, 2026, in preparation); until it has a DOI, cite the family's own source paper and the dmipy-sim commit recorded in the family's manifest.

models 0

None public yet