Spatiotemporal learning improves quantification of pore-space evolution and fluid displacement
Authors: Zhuangzhuang Ma, Rukuan Chai, Qianqian Ma, Yanghua Wang, Branko Bijeljic, Martin J. Blunt
Organizations: Resource Geophysics Academy, Imperial College London, London, SW7 2BP United Kingdom. · Department of Earth Science and Engineering, Imperial College London, London, SW7 2AZ, United Kingdom.
Pore-space evolution and fluid displacement influence how rocks transmit and store fluids, affecting geological carbon storage, underground hydrogen storage and groundwater contaminant transport. A single time-resolved micro-CT experiment may produce tens to hundreds of three-dimensional images, and segmentation is required to quantify physical properties and run image-based simulations. Traditional methods such as thresholding perform poorly on noisy, low-contrast images and need manual adjustment per image; 3D deep learning networks segment each image without temporal information, so their errors vary between images and could be mistaken for physical change. Here we introduce a transformer-based 4D segmentation framework combining spatial and temporal information. It uses sequential images to resolve ambiguous phase assignments and separate physical change from segmentation error. The method is tested on images of mineral dissolution: it measures pore-volume growth and rock boundaries more accurately than an equivalent 3D model that ignores temporal information. It also captures the increase in permeability verified against reference segmentations and flow simulations. The advantage transfers to another dissolution sample without retraining. On two other datasets of fluid displacement, it reduces spurious changes in predicted rock and porosity fluctuations and accurately identifies filling events. Spatiotemporal segmentation turns sequences of hundreds of three-dimensional images into more reliable quantitative measurements of pore-space evolution and fluid displacement without per-image parameter adjustment to support the analysis of time-resolved CT imaging.
Figures & tables
Figure 1: A framework for measuring evolving pore geometry and fluid occupancy. a , Subsurface processes motivate time-resolved imaging of rock samples. b , Spatial attention combines spatial information within each image; temporal attention combines information from other images in the time sequence. c , Training with labels at selected time steps. Input corruption and a process-specific consistency loss are optional settings. d , Segmented geometry provides porosity and saturation measurements and domains for flow and reactive-transport simulations.
Figure 2: Measuring pore-space evolution during carbonate dissolution. a , Pore–rock surfaces at 300 min. Predicted surfaces are coloured by signed distance to the reference boundary: brown indicates locally smaller pores and green larger pores. b , A greyscale cross-section and segmentations of the boxed region. Purple marks disagreement between the predictions and the reference segmentation, including misclassification between oil and brine. c , Absolute porosity and oil-saturation errors, rock boundary F1 score, and fluid Dice score across all ten images with reference segmentations. At each time step, points show means and error bars show ±1 sample standard deviation across three training runs with different random seeds. Error bars smaller than their markers are not visible.
Figure 3: Effects of segmentation on simulated flow and mineral dissolution rate. a,b , Relative errors in permeability and total mineral dissolution rate versus relative porosity error, calculated as 100(X/Xref−1) , where Xref is obtained from the reference segmentation at the same time step. Each model point represents one training run. Random-perturbation points show means, with ranges in Supplementary Table 16 . Net-zero perturbations preserve porosity by converting equal numbers of rock voxels to pore and pore voxels to rock. c , Pressure averaged over pore voxels in each cross-section along the flow direction. The plotted interval (0.61–2.44 mm from the inlet) was used to calculate permeability. Pressure was set to zero at the outlet of the full simulation domain. d , Permeability divided by its value at 0 min, K(t)/K(0) . Model bars show means and error bars show ±1 sample standard deviation across three training runs with different random seeds. e,f , Velocity magnitude along streamlines and concentration at rock surfaces. Velocity is normalised by the mean pore velocity in each simulation; concentration by the inlet concentration. Velocity errors compare values at the same spatial position; concentration errors compare values at the nearest corresponding rock-surface points. Black surface regions have no reference surface within approximately 12 μ m and are excluded from the concentration-error comparison.
Figure 4: The contribution of other images in the time sequence to dissolution segmentation. a , Different model inputs for the prediction at 300 min. The 3D model was retrained with input corruption; the masking example is enlarged for display. The current-time-step-only 4D model processed each image independently during training and inference. Repeated-input 4D uses the trained 4D model with repeated copies of the image at the current time step during inference. b–d , Mean absolute porosity error, fluid Dice and rock boundary F1 scores, averaged over all ten time steps within each training run. Points show means and error bars show ±1 sample standard deviation across three training runs with different random seeds. Dotted lines mark the mean for the 3D model. Error bars smaller than their markers are not visible. e , Segmentation faces of the same 1283 -voxel region at 300 min. Purple marks misclassified voxels. f , Rock boundary F1 scores for the four time steps where the 3D mean is below 0.80 and for the remaining six.
Figure 5: Temporal consistency of the solid geometry and segmentation with limited training labels. a,b, Porosity deviation, Δϕ , in percentage points (pp), for Ketton drainage and imbibition (a) and Luxembourg sandstone (b). Medians use the complete Ketton sequences and time steps at 5–69 min for Luxembourg sandstone. The temporal median of each run was subtracted before averaging across training runs. The sandstone interval excludes sequence ends, where input windows repeat the image at the nearest endpoint. The dotted line in b marks oil arrival in the evaluation region. The sandstone panels b and c use models trained with one labelled training time step. c, Changes in predicted rock voxels between consecutive time steps (left), measured as the percentage of evaluated voxels whose predictions changed between rock and pore, and mean absolute porosity change (right). Statistics use pairs ending at time steps 26–471 for drainage and 26–391 for imbibition. The later time step in each sandstone pair is at 5–69 min, including the 4–5-min pair. Printed percentages give the relative change of the 4D mean from the 3D mean. d, Luxembourg sandstone brine Dice score and mean absolute oil-saturation error with one, three or five labelled training time steps, averaged over the designated evaluation times, excluding the 37-min time step. Blue and red denote brine and oil; hollow circles with dashed lines denote 3D, and filled squares with solid lines denote 4D. Curves and bars show means over three training seeds; shading and error bars indicate ±1 sample standard deviation.
Figure 6: Fluid occupancy changes and their recovery from dynamic images. a, Luxembourg sandstone oil-filling volumes: connected brine-to-oil changes over sampled 5- or 6-min intervals. Time steps give scan start times relative to continuous imaging. b, Ketton limestone total and net change: the sum and difference, respectively, of brine-to-oil and oil-to-brine transition volumes over consecutive time steps. Volumes were summed over consecutive image pairs within each training run and then averaged across the three runs. c, Brine and oil Dice scores at common evaluation time steps when the input images were separated by 60, 180 or 300 s. Squares denote 4D; dashed lines denote 3D. d, During the scan at 35 min, a filling pore in the sandstone resembles rock because of intermediate greyscale values. The 3D prediction assigns it to rock; the reference segmentation and 4D prediction assign it to brine. Magenta outlines the reference pore. e, The largest reference filling region, at 32–37 min. Red and grey show its volume captured and missed by predicted brine-to-oil changes. Renderings show one seed; percentages average three. Means and shading or error bars represent three seeds and ±1 sample standard deviation. Filtering and recovery calculations are described in Supplementary Note Supplementary Note 7. Measuring oil filling between time steps .
Seq2 (mixed-wet)
Seq3 (water-wet)
Metric
3D
4D
3D
4D
Rock Dice score
0.9672±0.0006
0.9763±0.0007
0.9513±0.0024
0.9555±0.0005
Oil Dice score
0.8258±0.0045
0.8275±0.0029
0.8516±0.0023
0.8678±0.0029
Brine Dice score
0.8709±0.0036
0.9042±0.0036
0.7552±0.0070
0.7745±0.0011
Fluid Dice score
0.8484±0.0007
0.8658±0.0027
0.8034±0.0030
0.8212±0.0017
Rock BF1 score
0.4637±0.0103
0.6087±0.0038
0.3661±0.0151
0.3803±0.0009
Table 1: Segmentation transfer without fine-tuning. Seq1-trained 3D and 4D models were applied to different experiments Seq2 and Seq3 without target adaptation or target-based checkpoint selection. Scores cover the full spatial extent of the reference segmentations within valid sample support. Values are means ± sample standard deviations across three seeds, after averaging the two time steps with reference segmentations within each seed. Fluid scores equally weight oil and brine. Unlike the central time-step output used for Seq1, 4D transfer predictions were combined across overlapping temporal windows using Gaussian weights.
Dataset / process
Imaging technique
Voxel size ( μ m)
Data source
Ketton limestone: water-wet dissolution (Seq1)
Laboratory micro-CT
6.09
Ma et al. (2026)
Ketton limestone: mixed-wet dissolution (Seq2)
Laboratory micro-CT
5.2
Ma et al. (2026)
Ketton limestone: water-wet dissolution (Seq3)
Laboratory micro-CT
5.2
Ma et al. (2026)
Ketton limestone: drainage
Synchrotron micro-CT
3.28
Singh et al. (2018)
Ketton limestone: imbibition
Synchrotron micro-CT
3.28
Singh et al. (2018)
Luxembourg sandstone: oil flooding
Laboratory micro-CT
8.051
Mascini et al. (2021) ; Mascini et al. (2021)
Table 2: Overview of the time-resolved micro-CT datasets. Voxel sizes are those used in this study.
Supplementary Fig. 1: Network features at phase boundaries during dissolution. Samples cover the complete 5003 Seq1 evaluation region, with identical coordinates for 3D and 4D. Features were extracted from 6283 inputs using 1283 patches, 50% overlap and Gaussian weighting ( σ=16 voxels), then L2-normalised. a , t-SNE projections for training seed 42, coloured by reference phase. Projections were fitted independently, so their coordinate axes cannot be compared. b , Phase purity in the original 48-dimensional feature space, with one point per time step and training seed (ten time steps and three seeds). c , Paired 4D-minus-3D differences in percentage points. The vertical axis uses a signed square-root scale, with ticks labelled in the original units. Boxes show the median and interquartile range; whiskers span the minimum to maximum of the 30 values. Mean brine–rock purity increased from 0.668 to 0.785. The paired increase was 11.72±0.76 percentage points, calculated by averaging over time steps within each seed and then taking the mean and sample standard deviation across three seeds.
Supplementary Fig. 2: Predicted phases in pores during oil filling at different greyscale values. The analysis includes 1,603 fixed reference pores of at least 100 voxels, each with one brine-to-oil transition between neibourghing time steps during 5–69 min in the combined validation and test region. The selection uses reference segmentations only. a , Fraction of each pore predicted as rock by the 3D model (circles) and the 4D model (squares) in the image preceding the reference transition, plotted against its relative greyscale position x . Values near 1 resemble the earlier brine-filled state; values near 0 resemble the later oil-filled state. The dashed line marks the median rock position ( xrock≈0.83 ). b , Volume-weighted rock fractions in the five greyscale intervals; n gives the number of pores. The bracket marks x<0.8 , the criterion used to identify filling during the scan. For 3D and 4D, values average training seeds 42, 1234 and 2026. Temporal voting uses 3D seed 42, with two preceding and two following time steps around the image being classified, including the central image. Definitions and per-seed results are given in Supplementary Note Supplementary Note 8. Predicted phases in pores during oil filling and Table 10 .
Supplementary Fig. 3: Detailed PoreTimeNet architecture. The model receives a temporal window of 3D images as input and predicts phase logits for all time steps within the window. A shared Swin-Transformer encoder extracts hierarchical spatial features from each 3D image, after which temporal self-attention propagates information across corresponding spatial locations in the temporal window. The decoder uses skip connections and decoder-stage temporal fusion to combine multi-scale spatial and temporal features before projecting the output to oil, brine and rock logits.
Metric
3D
3D + input corruption
Current-time-step-only 4D
4D
Rock Dice score
0.9833±0.0002
0.9831±0.0005
0.9830±0.0002
0.9910±0.0004
Fluid Dice score
0.9359±0.0008
0.9355±0.0016
0.9347±0.0002
0.9573±0.0016
Rock BF1 score
0.8334±0.0028
0.8306±0.0058
0.8294±0.0024
0.9408±0.0052
Fluid BF1 score
0.8594±0.0023
0.8573±0.0045
0.8567±0.0009
0.9313±0.0046
∣Δϕ∣ (pp)
0.9667±0.0991
0.9696±0.1118
1.0831±0.0876
0.2786±0.0690
∣ΔSo∣ (pp)
1.3549±0.1542
1.2151±0.3309
1.5032±0.0577
0.9411±0.1924
Supplementary Table 1: Segmentation accuracy during dissolution. The boundary F1 (BF1) score measures agreement with reference boundaries at Euclidean distances d≤1 voxel. ∣Δϕ∣ and ∣ΔSo∣ denote absolute porosity and oil-saturation errors in percentage points (pp). Values are means ± sample standard deviations across three training seeds after averaging within each seed. Fluid Dice and BF1 scores are equally-weighted means over oil and brine. The current-time-step-only 4D model processed each image independently during training and inference.
Model
Pore-to-rock (%)
Fluid Dice score
3D
1.6819±0.0170
0.9359±0.0008
3D + input corruption
1.6766±0.0335
0.9355±0.0016
3D + voting
0.5549±0.0095
0.8990±0.0007
4D
1.0441±0.0103
0.9573±0.0016
Repeated-input 4D
2.0691±0.0342
0.9216±0.0025
Current-time-step-only 4D
1.6653±0.0124
0.9347±0.0002
Supplementary Table 2: Pore-to-rock changes in predictions and segmentation accuracy. Pore-to-rock values are percentages of voxels in the complete evaluation region per interval. Values are means ± sample standard deviations across three training seeds after averaging within each seed. The fluid Dice score is the equally-weighted mean of oil and brine Dice scores. Voting uses three consecutive time steps; current-time-step-only 4D and Repeated-input 4D are defined in main-text Fig. 4 a.
Metric
3D
3D + input corruption
4D
Rock Dice score
0.9938±0.0002
0.9940±0.0001
0.9941±0.0002
Fluid Dice score
0.9065±0.0001
0.9080±0.0001
0.9088±0.0008
Rock BF1 score
0.9352±0.0043
0.9415±0.0022
0.9422±0.0037
Fluid BF1 score
0.9219±0.0004
0.9232±0.0014
0.9247±0.0008
∣Δϕ∣ (pp)
0.3194±0.1160
0.3114±0.0842
0.3172±0.0886
∣ΔSo∣ (pp)
0.6955±0.4100
0.7189±0.3462
0.7027±0.2033
Supplementary Table 3: Segmentation accuracy and consistency during Ketton drainage. Changes in predicted rock between consecutive time steps were averaged over time steps 26–471, excluding 25 time steps at each end of the sequence. Mean absolute porosity change is the mean of the absolute differences between consecutive time steps. Values are means ± sample standard deviations across three training seeds. ∣Δϕ∣ and ∣ΔSo∣ denote absolute porosity and oil-saturation errors in percentage points (pp). BF1 denotes the boundary F1 score with a one-voxel matching tolerance. Fluid Dice and Fluid BF1 scores average the corresponding oil and brine scores equally.
Metric
3D
3D + input corruption
4D
Rock Dice score
0.9953±0.0002
0.9954±0.000025
0.9957±0.0001
Fluid Dice score
0.9442±0.0013
0.9455±0.0002
0.9510±0.0006
Rock BF1 score
0.9810±0.0009
0.9805±0.0003
0.9842±0.0003
Fluid BF1 score
0.9490±0.0021
0.9499±0.0005
0.9567±0.0008
∣Δϕ∣ (pp)
0.1511±0.0446
0.1843±0.0252
0.1360±0.0108
∣ΔSo∣ (pp)
0.7096±0.1508
0.5427±0.0556
0.3263±0.1205
Supplementary Table 4: Segmentation accuracy and consistency during Ketton imbibition. Changes in predicted rock between consecutive time steps were averaged over time steps 26–391, excluding 25 time steps at each end of the sequence. Mean absolute porosity change is the mean of the absolute differences between consecutive time steps. Values are means ± sample standard deviations across three training seeds. ∣Δϕ∣ and ∣ΔSo∣ denote absolute porosity and oil-saturation errors in percentage points (pp). BF1 denotes the boundary F1 score with a one-voxel matching tolerance. Fluid Dice and Fluid BF1 scores average the corresponding oil and brine scores equally.
Process
Metric
3D
4D
Repeated-input 4D
Drainage
Fluid Dice score
0.9065±0.0001
0.9088±0.0008
0.9082±0.0001
Drainage
Changes in predicted rock (%)
0.5448±0.0136
0.3758±0.0182
0.5140±0.0063
Imbibition
Fluid Dice score
0.9442±0.0013
0.9510±0.0006
0.9494±0.0006
Imbibition
Changes in predicted rock voxels (%)
0.4904±0.0105
0.3214±0.0081
0.4828±0.0043
Supplementary Table 5: Ketton segmentation with repeated images. The fluid Dice score uses the same evaluation regions as the preceding drainage and imbibition tables. For the comparison of changes in predicted rock, we used predictions at time steps 219–278 for drainage and 179–238 for imbibition, and evaluated changes at time steps 221–278 and 181–238, respectively (58 scored time steps per process, after two inputs initialise consecutive-transition metrics). Changes in predicted rock voxels are quantified as the percentage of evaluated voxels that switch between rock and pore in consecutive predictions, averaged over these time steps. The same voxels were evaluated at all time steps within each process. Values are means ± sample standard deviations across three training seeds; these rates describe the selected time interval. The fluid Dice score is the equally-weighted mean of oil and brine Dice scores.
Number of labelled training time steps
Model
Brine Dice score
1
3D
0.6388±0.0018
1
4D
0.7127±0.0019
3
3D
0.6458±0.0012
3
4D
0.7056±0.0108
5
3D
0.6441±0.0009
5
4D
0.7088±0.0066
Supplementary Table 6: Effect of the number of labelled training time steps on segmentation performance. The brine Dice score uses the 14 evaluation images remaining after excluding the image at 37 min. Values are means ± sample standard deviations across three training seeds after averaging within each seed. Evaluation times are listed in Supplementary Note Supplementary Note 5. Effect of the number of labelled training time steps on segmentation .
Number of labelled training time steps
Model
Brine-volume error (%)
∣Δϕ∣ (pp)
∣ΔSo∣ (pp)
1
3D
12.63±1.13
0.68±0.09
2.91±0.08
1
4D
11.56±1.31
0.65±0.07
2.44±0.34
3
3D
11.40±3.22
0.63±0.21
2.53±0.53
3
4D
8.73±6.25
0.53±0.45
1.59±0.62
5
3D
12.85±1.75
0.76±0.14
2.50±0.12
5
4D
10.12±1.46
0.59±0.11
2.00±0.18
Supplementary Table 7: Errors in porosity, oil saturation and brine volume. Brine error is the absolute relative volume error in percent. Porosity and oil-saturation errors are mean absolute errors in percentage points (pp), averaged over the retained evaluation times. Oil saturation was calculated using the pore volume in each segmentation. Values are means ± sample standard deviations across three training seeds after averaging within each seed. ∣Δϕ∣ and ∣ΔSo∣ denote the porosity and oil-saturation errors.
Number of labelled training time steps
Model and interval
Brine Dice score
Oil Dice score
Fluid Dice score
1
3D
0.6409±0.0018
0.7766±0.0133
0.7087±0.0068
1
4D, 60 s
0.7258±0.0027
0.8570±0.0049
0.7914±0.0016
1
4D, 180 s
0.7340±0.0023
0.8577±0.0034
0.7958±0.0013
1
4D, 300 s
0.7402±0.0018
0.8631±0.0040
0.8017±0.0012
3
3D
0.6477±0.0012
0.7265±0.0200
0.6871±0.0102
3
4D, 60 s
0.7159±0.0112
0.8586±0.0019
0.7873±0.0058
Supplementary Table 8: Segmentation accuracy at different time intervals between input images. The fluid Dice score is the equally-weighted mean of oil and brine Dice scores, calculated within each seed. Scores use the same 11 evaluation time points listed in Supplementary Note Supplementary Note 6. Effect of the time interval between input images on segmentation accuracy . Values are means ± sample standard deviations across three training seeds after averaging over time within each seed. The 3D predictions are reused at every input interval. The rows with one labelled training time step correspond to main-text Fig. 6 c.
Segmentation
Volume (mm 3 )
Difference (%)
Reference segmentation
1.942
—
3D
1.095±0.009
−43.61±0.46
4D
1.658±0.002
−14.62±0.13
Supplementary Table 9: Oil-filling volume across the sampled intervals. For the Luxembourg sandstone models trained with one labelled time step, filling denotes brine-to-oil transitions over the sampled intervals (Supplementary Note Supplementary Note 7. Measuring oil filling between time steps ). Model values are means ± sample standard deviations across three training seeds; reference volume is common to all comparisons. Signed difference is 100(Vmodel−Vreference)/Vreference .
3D
3D vote
4D
Subset
n
42
1234
2026
42
42
1234
2026
All pores
1603
0.708
0.697
0.695
0.721
0.172
0.187
0.181
x<0.8
922
0.909
0.891
0.893
0.908
0.185
0.200
0.191
0.6≤x<0.8
496
0.968
0.959
0.961
0.967
0.184
0.203
0.191
x≥0.8
681
0.469
0.465
0.460
0.498
0.157
0.172
0.170
Supplementary Table 10: Pore volume predicted as rock before reference oil filling. Values are volume-weighted rock fractions at t−1 within each pore subset. The 3D and 4D columns show individual training seeds. Temporal voting uses only 3D seed 42 and a five-image window centred on t−1 ; it is not a mean over seeds. The relative greyscale position x and the event selection are defined in Supplementary Note Supplementary Note 8. Predicted phases in pores during oil filling .
Metric
3D
nnU-Net
4D
Rock Dice score
0.9833±0.0002
0.9800
0.9910±0.0004
Fluid Dice score
0.9359±0.0008
0.9223
0.9573±0.0016
Rock BF1 score
0.8334±0.0028
0.7920
0.9408±0.0052
Fluid BF1 score
0.8594±0.0023
0.8061
0.9313±0.0046
∣Δϕ∣ (pp)
0.97±0.10
1.11
0.28±0.07
∣ΔSo∣ (pp)
1.35±0.15
1.49
0.94±0.19
Supplementary Table 11: Comparison with nnU-Net. Segmentation scores cover all ten time steps and the complete 5003 region. For these scores, 3D and 4D values are means ± sample standard deviations across three training runs after averaging over time within each run; nnU-Net values are time averages from one run. Porosity and saturation errors are averaged after taking their absolute values. Fluid Dice and Fluid BF1 scores average oil and brine equally. The BF1 score uses a one-voxel matching tolerance. ∣Δϕ∣ and ∣ΔSo∣ are absolute porosity and oil-saturation errors in percentage points (pp). Permeability ( K ) and total mineral dissolution rate ( R ) are evaluated at the stated times using the same flow and reactive-transport settings as main-text Fig. 3 : K is calculated from the pressure gradient fitted over planes 100–399. For the K and R error rows, 3D and 4D values are means ± sample standard deviations of the three absolute relative errors, whereas nnU-Net values are signed relative errors from one run; positive and negative signs denote overestimation and underestimation, respectively. The final row reports the permeability ratio, whose reference value is 8.34. No between-run standard deviation is available for nnU-Net.
Setting
Dissolution
Ketton displacement
Luxembourg sandstone
Spatial patch size (voxels)
1283
963
963
Images per input window
5 images
11 images
11 images
Training input intervals
Consecutive time steps
1, 3, 5, 10 time-step intervals
1, 3, 5 time-step intervals
Epochs / warm-up
250 / 50
300 / 30
300 / 30
Peak learning rate
2×10−5
1.4142×10−5
1.4142×10−5
Labelled training time steps
10
6 per process
1, 3 or 5
Supplementary Table 12: Training and inference settings for each dataset. An input window contains three-dimensional patches from the same spatial region at several time steps. The 3D model processes the images independently; the 4D model combines their information. Intervals below count positions in the image sequence. The models started from random initialisation and used two GPUs with one window per GPU and no gradient accumulation. AdamW weight decay was 10−5 . We retained the checkpoint with the highest validation fluid Dice score. Comparisons with input corruption, current-time-step-only 4D and repeated inputs are described separately.
Time (min)
Model
K
R
δK (%)
δR (%)
0
3D
7.765±0.663
3.193±0.113
+12.99±9.64
−4.15±3.39
3D + input corruption
7.819±1.128
3.270±0.067
+13.77±16.42
−1.82±2.02
4D
6.723±0.354
3.180±0.041
−2.18±5.15
−4.53±1.24
187
3D
30.612±1.020
3.769±0.023
−6.71±3.11
−3.13±0.59
3D + input corruption
30.239±1.890
3.759±0.036
−7.85±5.76
−3.39±0.92
4D
32.802±0.223
3.838±0.025
−0.03±0.68
−1.34±0.65
Supplementary Table 13: Permeability and total mineral dissolution rate. Values are means ± sample standard deviations across three training seeds (42, 1234 and 2026) at each time. K is in 10−12m2 and R in 10−9mols−1 . Signed errors are calculated for each seed as 100(X/Xref−1) before summarising across seeds; they retain the direction of the deviation from the reference. The standard deviations describe variation between training runs on the same sample. All simulations use the native voxel grid.
Interval (min)
Segmentation
K(t)/K(0)
Absolute error (%)
0–187
Reference
4.77
—
3D
3.95±0.22
17.17±4.61
3D + input corruption
3.90±0.34
18.34±7.08
4D
4.89±0.25
4.78±1.03
0–646
Reference
8.34
—
3D
6.07±0.31
27.19±3.77
Supplementary Table 14: Permeability relative to 0 min. The ratio K(t)/K(0) gives permeability at time t relative to its initial value. Ratios are calculated within each seed before taking the mean and sample standard deviation across three training seeds. The final column is the mean and sample standard deviation of each seed's absolute relative deviation from the reference ratio.
Field
Pressure target
Velocity target
Velocity relaxation
0 min, reference
10−7
10−7
0.99
0 min, 3D
10−6
10−4
0.90
0 min, 4D
5×10−7
2×10−5
0.99
187 and 646 min, all groups
3×10−6
10−4
0.90
Supplementary Table 15: Flow-solver settings. These settings apply to the fields used for visualisation. Pressure and velocity targets are the outer-iteration initial residual thresholds; velocity relaxation is dimensionless. The velocity target applies to each component. Reference denotes the reference segmentation, and 3D and 4D denote the matched models at training seed 42.
Time (min)
Pore-volume change (%)
n
Permeability error range (%)
Total mineral dissolution rate error range (%)
0
−3
3
[−22.650,−22.420]
[12.132,12.362]
0
0
2
[−21.328,−21.039]
[26.920,26.955]
0
+3
3
[4.307,4.342]
[14.787,14.874]
187
−3
3
[−18.836,−18.790]
[15.992,16.016]
187
0
2
[−18.084,−17.852]
[33.901,33.904]
187
+3
3
[3.800,3.834]
[17.340,17.374]
Supplementary Table 16: Variation across repeated random perturbations of pore boundaries. Each range gives the minimum and maximum relative error across n random draws at the same time step, using the reference calculation at that time. Main-text Fig. 3 a,b shows the corresponding means. A zero pore-volume change exchanges 3% of the original pore-voxel count in each direction. Only settings with repeated draws are listed; all other settings used one draw.
Multiphase flow in porous media underpins subsurface energy and environmental technologies, including geological CO2 storage and underground hydrogen storage, yet pore-scale dynamics in realistic three-dimensional materials remain difficult to characterize and predict. Here we introduce a multimodal learning framework that infers multiphase pore-scale flow directly from time-resolved four-dimensional (4D) micro-velocimetry measurements. The model couples a graph network simulator for Lagrangian tracer-particle motion with a 3D U-Net for voxelized interface evolution. The imaged pore geometry serves as a boundary constraint to the flow velocity and the multiphase interface predictions, which are coupled and updated iteratively at each time step. Trained autoregressively on experimental sequences in capillary-dominated conditions (Ca≈10−6), the learned surrogate captures transient, nonlocal flow perturbations and abrupt interface rearrangements (Haines jumps) over rollouts spanning seconds of physical time, while reducing hour-to-day--scale direct numerical simulations to seconds of inference. By providing rapid, experimentally informed predictions, the framework opens a route to ''digital experiments'' to replicate pore-scale physics observed in multiphase flow experiments, offering an efficient tool for exploring injection conditions and pore-geometry effects relevant to subsurface carbon and hydrogen storage.
Chunyang Wang, Linqi Zhu, Yuxuan Gu +9
Department of Earth Science and Engineering, Imperial College London. · Department of Geology, Ghent University. · Department of Energy Science and Engineering, Stanford University. +2
Accurate prediction of permeability tensors from pore-scale microstructure images is essential for subsurface flow modeling, yet direct numerical simulation requires hours per sample, fundamentally limiting large-scale uncertainty quantification and reservoir optimization workflows. A physics-informed deep learning framework is presented that resolves this bottleneck by combining a MaxViT hybrid CNN-Transformer architecture with progressive transfer learning and differentiable physical constraints. MaxViT's multi-axis attention mechanism simultaneously resolves grain-scale pore-throat geometry via block-local operations and REV-scale connectivity statistics through grid-global operations, providing the spatial hierarchy that permeability tensor prediction physically requires. Training on 20000 synthetic porous media samples spanning three orders of magnitude in permeability, a three-phase progressive curriculum advances from an ImageNet-pretrained baseline with D4-equivariant augmentation and tensor transformation, through component-weighted loss prioritizing off-diagonal coupling, to frozen-backbone transfer learning with porosity conditioning via Feature-wise Linear Modulation (FiLM). Onsager reciprocity and positive definiteness are enforced via differentiable penalty terms. On a held-out test set of 4000 samples, the framework achieves variance-weighted R2 = 0.9960 (R2_Kxx = 0.9967, R2_Kxy = 0.9758), a 33% reduction in unexplained variance over the supervised baseline. The results offer three transferable principles for physics-informed scientific machine learning: large-scale visual pretraining transfers effectively across domain boundaries; physical constraints are most robustly integrated as differentiable architectural components; and progressive training guided by diagnostic failure-mode analysis enables unambiguous attribution of performance gains across methodological stages.
Mohammad Nooraiepour
Faculty of Mathematics and Natural Sciences, University of Oslo, PO Box 1047 Blindern, 0316 Oslo, Norway
Carbonate reservoirs offer significant capacity for subsurface carbon storage, oil production, and underground hydrogen storage. X-ray computed tomography (X-ray CT) coupled with numerical simulations is commonly used to investigate the multiphase flow behaviors in carbonate rocks. Carbonates exhibit pore size distribution across scales, hindering the comprehensive investigation with conventional X-ray CT images. Imaging samples at both macro and micro-scales (multi-scale imaging) proved to be a viable option in this context. However, multi-scale imaging faces two key limitations: the trade-off between field of view and voxel size necessitates resource-intensive imaging, while multi-scale multi-physics numerical simulations on resulting digital models incur prohibitive computational costs. To address these challenges, we propose a machine learning-enhanced data assimilation framework that leverages experimental drainage relative permeability measurements to achieve efficient characterization of micro-scale structures, delivering a data-driven solution toward a high-fidelity multiscale digital rock modeling. We train a dense neural network (DNN) as a proxy to a multi-scale pore network simulator and couple it with an ensemble smoother with multiple data assimilation (ESMDA) algorithm. DNN-ESMDA framework simultaneously infers the CO2-brine drainage relative permeability of microporosity phases with associated uncertainty estimation, revealing the relative importance of each rock phase and guiding future characterization. Our DNN-ESMDA framework achieves a computational speedup, reducing inference time from thousands of hours to seconds compared with the usage of conventional multiscale numerical simulation. Given this computational efficiency and applicability, the machine learning-enhanced ESMDA framework presents a generalizable approach for characterizing multiscale carbonate rocks.
Zhenkai Bo, Ahmed H. Elsheikh, Hannah P. Menke +5
Institute of GeoEnergy Engineering, Heriot-Watt University, Edinburgh, EH14 4AS, United Kingdom · Department of Geoscience & Engineering, Technical University of Delft, Delft, The Netherlands · PETRONAS Research Sdn Bhd, Bangi, Selangor, Malaysia