ling-sand-packs

Resin-set quartz and garnet sand packs whose CPMG was measured and whose pore space was then imaged: the relaxometry reference where the measurement and the geometry are the SAME object. Relaxation is not in the walk, so one pack answers every relaxivity. The reference it reproduces is Model Synthetic Samples for Validation of NMR Signal Simulations (Transport in Porous Media), on the same object: grade A by the rule of dmipy-sim#459. These packs disagree with the published number. 1_G100: replayed 222.049 against 178.449 printed in Model Synthetic Samples for Validation of NMR Signal Simulations (Ling et al. 2022), whose released te = 100 us echo train the value is read from, the log-mean of the released te = 100 us CPMG train of sample 1_G100, inverted on the 1 ms grid with the same estimator as ours: 24.433% of it, 6.17 standard uncertainties. The publication states no uncertainty (none stated: the paper prints no T2 value anywhere -- its comparison is Fig. 5, two distributions overlaid with no agreement metric -- and the released echo train carries no repeat, so the gate's budget rests on our own standard errors alone and the disagreement with them is recorded rather than failed on), so this disagreement cannot be failed on and is recorded instead. 6_Q100: replayed 855.685 against 727.084 printed in Model Synthetic Samples for Validation of NMR Signal Simulations (Ling et al. 2022), whose released te = 100 us echo train the value is read from, the log-mean of the released te = 100 us CPMG train of sample 6_Q100, inverted on the 1 ms grid with the same estimator as ours: 17.687% of it, 4.49 standard uncertainties. The publication states no uncertainty (none stated: the paper prints no T2 value anywhere -- its comparison is Fig. 5, two distributions overlaid with no agreement metric -- and the released echo train carries no repeat, so the gate's budget rests on our own standard errors alone and the disagreement with them is recorded rather than failed on), so this disagreement cannot be failed on and is recorded instead. No gate check fails on this: the publication states no uncertainty to fail it against, so the disagreement is a recorded number rather than a verdict. 2 of the gate's 60 checks failed; § The gate below names each one.

Every number on this card is read from a record in records/ of this dataset, written by the stage that measured it (the reference-pack protocol, dmipy-sim#482); none of it is transcribed, so this card, the gate and any paper read the same files.

The packs

The trade the design records: the target floor 0.005 needs 347,438 walkers and the 60 GB memory budget allows 115,564 (measured: 0.51 MB resident per walker in the pack stage on this window). The design therefore sets 115,564 walkers; the tier that holds its target at that count is contact, and every tier's achieved floor is certified in the pack and shown on the card.

The save grid is 1000 us over 3.5 s (3,501 saves): 1 ms is the grid the log-mean is stable on (measured on the Imperial rocks: a 20 ms sampling moved it by a factor of two) and it holds every 100 us echo of their train within a save. Each PUBLISHED pack's own window is its released train's extent, which is why the five do not share one and each is in the walk record

• the pack was walked at 130,000 walkers where this design's pilot sets 115,564; the walk is not repeated, and what holds it to the budget is its own RECORDED peak of 42.8 GB rather than the pilot's projection

What is inside

Three orthogonal cross-sections through the centre of each substrate, rendered from the spec the pack embeds by dmipy_sim.spec.preview -- the same membership test the walk uses, so the picture cannot show a substrate the walk does not have.

1_G100

• surface kinds: label_volume; pools free, garnet
• pixel 3.935 um, scale bar 200 um
• area fraction in the x-centre section: free 0.3899, garnet 0.6101 (the spec's realisation: porosity 0.3989, surface_to_volume 7.196e+04)

2_Q25G75

• surface kinds: label_volume; pools free, quartz, garnet
• pixel 3.936 um, scale bar 200 um
• area fraction in the x-centre section: free 0.3917, garnet 0.5007, quartz 0.1077 (the spec's realisation: porosity 0.3834, surface_to_volume 7.445e+04)

4_Q50G50

• surface kinds: label_volume; pools free, quartz, garnet
• pixel 3.93 um, scale bar 200 um
• area fraction in the x-centre section: free 0.3946, garnet 0.3226, quartz 0.2828 (the spec's realisation: porosity 0.3973, surface_to_volume 7.186e+04)

5_Q75G25

• surface kinds: label_volume; pools free, quartz, garnet
• pixel 3.936 um, scale bar 200 um
• area fraction in the x-centre section: free 0.3777, garnet 0.1150, quartz 0.5072 (the spec's realisation: porosity 0.3821, surface_to_volume 7.068e+04)

6_Q100

• surface kinds: label_volume; pools free, quartz
• pixel 3.935 um, scale bar 200 um
• area fraction in the x-centre section: free 0.3762, quartz 0.6238 (the spec's realisation: porosity 0.389, surface_to_volume 7.661e+04)

The reproduction

Grade A. the released data itself on the same released geometry: no free parameter marked 'ours' changes it. Their sample: the same physical object: the CPMG was measured on each pack, which was then resin-set and micro-CT imaged, and both halves are in the same deposit (the same object).

The quantity is compared on the grid the reference record states, with the solver it states; the gate refuses a reproduction on any other grid, because the grid is part of the measurement.

The free parameters, and whose they are:

• rho = {'garnet': 9.85e-05, 'quartz': 1.25e-05} m/s -- theirs (paper §3.2): fitted by them per mineral to their own te = 100 us trains; an effective value tied to their 3.93 um voxel, because a segmentation's surface is the Manhattan area at that resolution, which is what both walks relax against
• D0 = 2.3e-09 m^2/s -- theirs (paper §2.3): the brine's self-diffusion, stated
• T2B = 3.0 s -- ours (not in the paper): the paper states no bulk T2; 3.0 s is ours, applied at replay and not in the walk, and the reproduction reports the log-mean over 2.4-3.1 s as well
• T2 inversion = non-negative least squares with a second-difference penalty, lam = 0.1 (t2_distribution) - -- ours (the paper states its own only as a regularised inversion): non-negative least squares with a second-difference penalty on 60 points from 1 ms, applied to THEIR released train and to ours identically, so the two sides of the comparison share the estimator even though it is not theirs
• surface rule = exp(-2 (rho/D) d_perp) per reflection - -- ours (the walk): the boundary local time is stored as a channel and weighted at replay; their walk kills a walker at an attempted move into a grain voxel. The two agree in the continuum limit

Where each number comes from:

1_G100 -- Model Synthetic Samples for Validation of NMR Signal Simulations (Ling et al. 2022), whose released te = 100 us echo train the value is read from, the log-mean of the released te = 100 us CPMG train of sample 1_G100, inverted on the 1 ms grid with the same estimator as ours (10.1007/s11242-022-01764-w, resolved via crossref at 2026-09-27T01:17:40Z as 'Model Synthetic Samples for Validation of NMR Signal Simulations'): read from https://doi.org/10.6084/m9.figshare.17161730, sha256 0eb1de561db7

6_Q100 -- Model Synthetic Samples for Validation of NMR Signal Simulations (Ling et al. 2022), whose released te = 100 us echo train the value is read from, the log-mean of the released te = 100 us CPMG train of sample 6_Q100, inverted on the 1 ms grid with the same estimator as ours (10.1007/s11242-022-01764-w, resolved via crossref at 2026-09-27T01:17:40Z as 'Model Synthetic Samples for Validation of NMR Signal Simulations'): read from https://doi.org/10.6084/m9.figshare.17161730, sha256 a9e09a21776d

Caveats:

• echo_spacing -- Only their te = 100 us train is reproduced: it is the one their relaxivities were fitted on. The te-DEPENDENCE of their trains (100 us to 20 ms) IS the garnet's internal-gradient effect, and a label-volume pack carries no C3 field channel, so a longer-te train is refused by name rather than served as a decay without its dephasing.
• grade -- The pre-protocol record of this family stated grade B as a FIELD, reasoning that the paper prints no T2 value. grade_of assigns A by rule: the quantity is read from their own RELEASED echo trains, which ARE the measurement, on the same physical object. What the absence of a printed number costs is the reference's uncertainty -- it is 0.0 with a reason, so the published- comparison cannot be failed on and its disagreement is recorded instead.
• mixtures -- The three MIXTURES have no reference row. LabelVolume accumulates ONE boundary local time over every wall and Tissue.rho is one scalar, so a single pack cannot carry Ling's 12.5 um/s on quartz and 98.5 um/s on garnet at once. They are built, certified, gated and published -- the geometry, the positions and the contact channel are all sound -- and their spec declares NO nominal relaxivity, so a consumer must state one. Recorded as blocked (dmipy-sim#491), not approximated with one of the two numbers.
• windows -- The five packs do not share a window: 3.5 s for the two quartz-rich packs (35,000 echoes) and 2.0 s for the three garnet-rich ones (20,000 echoes), each the extent of ITS OWN released train. The design's window is the longest and each pack's own is in the walk record.

The gate

Deterministic, reading only the records: 60 checks, 2 failed. The tolerance is the design record's terms in quadrature -- an analytic standard error, the reference's own stated uncertainty, and any measured systematic -- with no coverage factor and no resampled error bar.

The failures, as the gate states them:

• 2_Q25G75/tier-contact-holds-as-designed: floor 0.00548 and codec error 2.84e-10 against the target 0.005; the design record predicted this tier would hold at 115,564 walkers and it does NOT: the pilot's scaling was falsified by the pack's own certificate
• 4_Q50G50/tier-contact-holds-as-designed: floor 0.00517 and codec error 2.46e-10 against the target 0.005; the design record predicted this tier would hold at 115,564 walkers and it does NOT: the pilot's scaling was falsified by the pack's own certificate

Use me

One call that reproduces one number of the table above. It was EXECUTED when this card was built (7.2 s, ceiling 60 s), against the local file of the same sha256 as packs/1_g100.rpk, and printed:

T2 log-mean 222.0 ms

import numpy as np
from dmipy_sim.replay import ReplayPack
from examples.validation.talabi_micro_ct_rocks import t2_distribution, log_mean_T2

pk = ReplayPack.load("hf://SubstrateCommons/ling-sand-packs/packs/1_g100.rpk")
ell = pk.contact()                                        # the stored wall-contact channel, per save
L = np.cumsum(np.asarray(ell, np.float64), axis=1)        # the boundary local time
t = np.arange(pk.n_t) * pk.dt
S = np.exp(pk.nominal.rho / 2.3e-09 * L - t / 3.0).mean(axis=0)   # pk.nominal.rho IS Ling's for this mineral
every = round(1.0e-3 / pk.dt)
g = np.logspace(-3, 1, 60)
print("T2 log-mean %.1f ms" % (1e3 * log_mean_T2(g, t2_distribution(t[every::every], S[every::every], g))))

Reference and licences

• the reference: Model Synthetic Samples for Validation of NMR Signal Simulations, https://doi.org/10.1007/s11242-022-01764-w -- cited, never redistributed. Ling et al. 2022 (Transport in Porous Media 142:623-639) for the packs, the measured CPMG and the fitted relaxivities; the micro-CT lattices and the echo trains are both in the figshare deposit 10.6084/m9.figshare.17161730 (CC BY 4.0).
• figshare-ling2022: https://doi.org/10.6084/m9.figshare.17161730 at figshare record 10.6084/m9.figshare.17161730 (2021, 349939116 bytes as the record states) -- CC-BY-4.0 (https://creativecommons.org/licenses/by/4.0/), the host's text copied verbatim to records/licences/figshare-ling2022-CC-BY-4.0.txt (18653 characters, sha256 9ba9550ad484); 6_Q100_Extracted_Voxel-3.93micro_X-450_Y-450_Z-450_Labels_0-Pore_1-Quartz_ASCII.am (a9e09a21776d, 273.4 MB), 5_Q75G25_Extracted_Voxel-3.93micro_X-450_Y-450_Z-450_Labels_0-Pore_1-Quartz_2-Garnet_ASCII.am (4a03008ce6d8, 273.4 MB), 4_Q50G50_Extracted_Voxel-3.93micro_X-450_Y-450_Z-450_Labels_0-Pore_1-Quartz_2-Garnet_ASCII.am (7f6e4b34431e, 273.4 MB), 2_Q25G75_Extracted_Voxel-3.93micro_X-450_Y-450_Z-450_Labels_0-Pore_1-Quartz_2-Garnet_ASCII.am (89d829221ba1, 273.5 MB), 1_G100_Extracted_Voxel-3.93micro_X-450_Y-450_Z-450_Labels_0-Pore_1-Garnet_ASCII.am (0eb1de561db7, 273.4 MB)
• the packs: CC-BY-4.0 images and echo trains; the packs are CC-BY-4.0 and no byte of the deposit is redistributed

Rendered by dmipy_sim.replay.reference from this dataset's records/. A bare republish of a pack regenerates the manifest-only card of dmipy_sim.replay.publish and drops this one; the protocol's publish stage writes this file last.