PoreML/PoreML_geo

PoreML-Geo

3D porous geometries for pore-scale drainage LBM simulation and ML surrogate modelling: ten rock/fiber subsets (synthetic + real micro-CT) and one designed flip-chip geometry, all with identical layout and deterministic seeded generation.

Conventions — arrays are boolean voxels with True = solid, axes ordered (z, y, x); fluid drains along the flow axis x (last), from x = 0 to x = −1. All samples percolate along the flow axis. Every size folder and asset is named by the saved array shape in that same (z, y, x) order — 128/256 for the cubic subsets, 256x256x128/128x128x64 for the anisotropic ones (so the trailing number is always the flow-axis length).

Quick start

from huggingface_hub import hf_hub_download
import numpy as np
solid = np.load(hf_hub_download("PoreML/PoreML_geo", "poly/128/poly128_0000.npy", repo_type="dataset"))

Subsets

Subset	Domain	N	Final porosity
poly/128/	128³	64	0.16 – 0.38
poly/256/	256³	64	0.21 – 0.31
blob/128/	128³	64	0.22 – 0.31
blob/256/	256³	64	0.22 – 0.31
sphere/128/	128³	64	0.17 – 0.33
sphere/256/	256³	64	0.16 – 0.33
bentheimer/128/	128³	64	0.17 – 0.42
bentheimer/256/	256³	64	0.19 – 0.29
castlegate/128/	128³	64	0.10 – 0.36
castlegate/256/	256³	64	0.18 – 0.26
buffberea/128/	128³	64	0.05 – 0.39
buffberea/256/	256³	64	0.10 – 0.28
gdl_ct/128x128x64/	128×128×64	64	0.48 – 0.67
gdl_ct/256x256x128/	256×256×128	64	0.53 – 0.62
gdl_ct_20/128x128x64/	128×128×64	64	0.43 – 0.68
gdl_ct_20/256x256x128/	256×256×128	64	0.53 – 0.63
gdl_ct_40/128x128x64/	128×128×64	64	0.52 – 0.68
gdl_ct_40/256x256x128/	256×256×128	64	0.56 – 0.63
fiber/128x128x64/	128×128×64	64	0.53 – 0.77
fiber/256x256x128/	256×256×128	64	0.53 – 0.77
flipchip/samples/	(gap+2)×482×(432–476)	64	see per-sample JSON (array region)

Geometry sources

Parameter sampling is deterministic (fixed seeds), and the realized generation parameters of every sample are recorded in its .json. Generation code is not included in this anonymized release.

• poly/ — synthetic polydisperse overlapping spheres (porespy, lognormal grain radii); poorly-sorted sandstone analogue.
• blob/ — thresholded Gaussian random fields (porespy blobs); tight/carbonate-like irregular pore morphology.
• sphere/ — monodisperse overlapping spheres at 2 μm/voxel; well-sorted sandstone as a zoomed-in high-resolution patch with wide throats.
• bentheimer/ — random crops of the DRP-317 Bentheimer sandstone micro-CT binary volume (1000³, 2.25 μm/voxel), augmented with random axis permutations + flips (full octahedral symmetry group).
• castlegate/ — same recipe applied to the DRP-317 CastleGate sandstone micro-CT binary volume (1000³, 2.25 μm/voxel; raw porosity ≈ 0.25).
• buffberea/ — same recipe applied to the DRP-317 Buff Berea sandstone micro-CT binary volume (1000³, 2.25 μm/voxel; raw porosity ≈ 0.23).
• gdl_ct/ — random crops of the real DRP-462 5% PTFE carbon-paper GDL micro-CT scan (Shojaei et al., Fuel 2023; 2.05 μm/voxel, 151 through-plane voxels of which 128 are cropped). No permutation/flip augmentation — one fixed transpose so the flow axis x is the GDL through-plane direction. Two sizes: the full 256×256×128 set and a half-linear-extent (1/8 volume) 128×128×64 set at the same 2.05 μm/voxel resolution and identical post-processing.
• gdl_ct_20/, gdl_ct_40/ — the same recipe (same seed-42 origin stream) applied to the DRP-462 20 % and 40 % PTFE-coating dry scans of the same GDL family (168 and 155 through-plane voxels; 1239² and 1284² in-plane inscribed-square parents, porosity ≈ 0.673 / 0.670). Together with gdl_ct (5 %) they form a PTFE-loading series at identical resolution and post-processing.
• fiber/ — synthetic in-plane fiber mats (porespy cylinders), the synthetic analog of gdl_ct: fibers lie in the transverse 256×256 (z, y) plane, isotropic in-plane with a small sampled out-of-plane tilt bound (phi_max 0–20°; fiber radius 3–5 vox); drainage is through-plane across the mat. Porosity-tolerance retry per sample. Two sizes matching gdl_ct: the full 256×256×128 set and a half-linear-extent (1/8 volume) 128×128×64 set at the same 1 μm/voxel resolution, fiber radii and porosity targets.
• flipchip/ — DESIGNED (non-rock) flip-chip underfill gaps for capillary underfill studies: a staggered array of truncated-sphere solder balls between a substrate and a die plate, compressed joint with real contact patches on both plates. Per sample, ball diameter D, standoff gap = squish·D (squish 0.60–0.95) and staggered pitch (1.5–2.1 D) are drawn deterministically, then every length is multiplied by a resolution scale of 1.5, giving the shipped D = 32–54 cells and gap ≥ 18 fluid layers; all balls share one voxel stencil, the array region is square, and balls keep ≥ 2 D clearance to the sealed y walls. x is open flow with 30-slab buffers and 1 D ball-free entry/vent margins.

Post-processing

All rock/fiber subsets: ball(3) morphological opening of the pore space (throats < 3 vox removed) + fill of face-disconnected pore pockets. flipchip is analytic — throats are designed (at the default 1.5 scale: ball–ball ≥ 15 cells at the equator, wall ≥ 9 cells) and recorded exactly in each JSON, so no morphological opening or SNOW2 extraction is applied.

Files per sample

Samples are named poly{size}_{NNNN}, blob{size}_{NNNN}, sphere{size}_{NNNN}, bent{size}_{NNNN}, castle{size}_{NNNN}, buff{size}_{NNNN}, gdl_{NNNN}, fiber_{NNNN}, flipchip_{NNNN}:

• .npy — boolean solid mask (see conventions above);
• .json — generation params + SNOW2 pore-network stats (flipchip: exact designed-geometry stats — ball centers, throats, contact-patch size);
• throat-distribution, cross-section and 3D-render figures with backing CSVs (flipchip: mid-gap top view instead of the throat histogram);
• per size folder: summary.csv.

Previews

Samples 0000–0015 of each subset; solid phase in 3D (left) and mid-plane slices with solid = gray, pore = white (right).

256³

	Solid phase (3D)	Mid-plane slices
poly
blob
sphere
bentheimer
castlegate
buffberea

128³

	Solid phase (3D)	Mid-plane slices
poly
blob
sphere
bentheimer
castlegate
buffberea

GDL, fiber & flipchip

	Solid phase (3D)	Mid slices / top views
gdl_ct 256×256×128
gdl_ct 128×128×64
gdl_ct_20 256×256×128
gdl_ct_20 128×128×64
gdl_ct_40 256×256×128
gdl_ct_40 128×128×64
fiber 256×256×128
fiber 128×128×64
flipchip

Pore & throat size distributions

SNOW2 inscribed radii aggregated over every sample in each subset; raw values back each figure in assets/{geom}_{size}_dist.csv.

	256³	128³
poly
blob
sphere
bentheimer
castlegate
buffberea

	256×256×128	128×128×64
gdl_ct
gdl_ct_20
gdl_ct_40
fiber

flipchip parameter coverage (ball diameter D vs gap height, colored by pitch):