When minerals crystallize inside rock, they can grow across the narrow necks between pores and seal them shut. This is why oil wells scale up, why injected CO₂ loses access to storage volume, and why chemical "scale inhibitors" are dosed into wells. Those inhibitors work by crystal growth poisoning: the inhibitor molecules stick to the sites where a crystal would otherwise add material, so the crystal stops growing. The mechanism is well documented in other settings. In dental plaque, for example, bisphosphonate and polyphosphate agents cut calcium deposition from 97.9 µg to 16.8 µg and 18.2 µg while leaving the organic scaffold untouched (Inhibitory effects of HEBP and Hexametaphosphate). What nobody can currently watch is which pore necks an inhibitor keeps open, and when the unprotected ones close, inside a real rock sample. Fast synchrotron micro-CT, which reconstructs 3D X-ray images from many angled projections, can image a core every 2 seconds at 2.75 µm (Learning Pore-scale Multi-phase Flow with GNN). To go faster or cut beam dose, however, one must take fewer projections per frame. The best sparse-view reconstructors then need a full iterative fit for every frame, and a 300-frame experiment turns into an offline backlog of hundreds of hours.
We propose a learned, gradient-free reconstructor that carries the 3D representation forward from one time frame to the next. The representation is a set of 3D Gaussian blobs, each with a position, a shape, and an X-ray attenuation value. For each new frame, a trained network reads the mismatch between the measured projections and the projections rendered from the previous frame's blobs. It then moves, reshapes, and re-weights existing blobs, and adds new ones where solid material has appeared. The claim is specific. On sequences where solid mineral grows into pore space, this recurrent update will recover pore-throat geometry and the timing of throat closure more accurately than a warm-started per-scan optimizer given the same wall-clock budget. It will do so at sub-second cost per frame and from roughly 10 to 30 projections per frame, few enough that the image sequence can distinguish inhibitor mechanisms that bulk chemistry cannot.
What exists and what is missing
Three lines of prior work each cover part of this problem. 3DGR-CT provides the representation and the physics. It models an attenuation volume as 3D Gaussians, renders projections with a differentiable CT projector, and initializes from a quick filtered back-projection (the classical one-shot CT reconstruction). It beats implicit neural fields and classical iterative solvers at 20 to 120 views. It is, however, fitted by gradient descent separately for every scan, and it has only been tested on static medical volumes. We reuse its projector and Gaussian parameterization but remove the per-scan fitting loop.
ReSplat supplies the update mechanism. In ordinary camera-based reconstruction, it renders the current Gaussians, computes 2D error maps against the input views, lifts those errors onto the Gaussians with attention, and predicts parameter offsets over four unrolled steps. The paper reports 100× faster reconstruction than per-scene 3D Gaussian Splatting and a 3.5 dB PSNR gain over DepthSplat on DL3DV 8-view synthesis. ReSplat operates on surface color with alpha blending. It starts every scene from scratch, and it never creates Gaussians where the scene has changed. We move its error-feedback idea to X-ray line integrals, whose physics follows the Beer-Lambert attenuation law, and we make it temporal. The single-pass feed-forward models GS-LRM and DepthSplat motivate the speed target, at 0.23 s and 0.6 s per scene respectively. They also constrain the design. Their depth-based and pixel-aligned placement assumes opaque surfaces and cannot see inside rock, and single-pass prediction tends to blur fine structure. For pore necks, blur is disastrous: capillary entry pressure scales as $1/r$ in the throat radius $r$, so smoothing a neck can open or close a capillary valve that controls trapping (Pore structure and desorption hysteresis in deformed coal).
On the rock side, two lines of work motivate the problem without solving it. LSE-MGN learns multiphase flow on a pore graph extracted once from a fully sampled scan. It assumes the solid never changes and needed 22.5 A100-hours of training for a small sub-volume. SPGAN and the MCMC reconstruction in LBM imbibition-flowback modeling generate statistically plausible rocks rather than the specific sample being altered, so they cannot follow one throat through time. What is missing from all of these is a reconstruction that is tied to the physical sample, handles a changing geometry, and is fast enough to keep up with the experiment.
Core Claim
Mineral precipitation and pore-throat closure have a structure that a learned temporal update can exploit and a warm-started optimizer cannot. The change between frames is sparse. It is roughly one-directional, since solid tends to grow into pores rather than recede. It is concentrated at solid–pore boundaries. And it requires new primitives in regions the previous frame left empty. Gradient descent warm-started from the previous frame must discover new material through the density-control heuristics of 3DGR-CT, which clone and split existing Gaussians. Under very sparse views it also tends to stay stuck near the previous solution. A network trained on growth sequences can learn where birth is likely from the back-projected residual and the local geometry, so it can place new solid in a few steps.
Research Program
The system keeps a Gaussian state $\mathcal{G}_{t-1}$ from the previous frame. At frame $t$ it receives $K_t$ projections $y_t$ at angles that rotate between frames, so that the angular coverage accumulates over time. Each of $S$ unrolled steps renders $A\,\mathcal{G}$ with the 3DGR-CT projector $A$, forms the residual $r = y_t - A\,\mathcal{G}$, and back-projects it into a coarse 3D residual volume. An attention module then gathers, for each Gaussian, the residual features along its rays together with those of its nearest neighbors, following ReSplat's error lifting. An update head predicts offsets to position, shape, and attenuation. A separate birth head proposes new Gaussians at voxels with high positive residual next to existing solid, and a pruning head removes Gaussians whose attenuation falls toward the pore value. Training minimizes projection error together with volume error against ground truth, plus a topology term that penalizes wrong throat radii. The throat radii are measured on a differentiable soft segmentation.
The training data comes from a simulator we build. Starting from real digital-rock volumes of Berea sandstone and Ketton limestone (SPGAN library) and from the synchrotron carbonate volumes of Spurin et al. (LSE-MGN data), we grow a mineral phase inward from the pore walls. Growth is supersaturation-driven and modulated by a spatial poisoning field. This field encodes two contrasting inhibitor mechanisms. In one, growth slows uniformly everywhere. In the other, crystal growth is blocked outright at a random subset of wall sites, which matches the site-adsorption picture in the HEBP study. Each synthetic sequence is rendered into sparse, noisy projections, with realistic photon noise and slight sample motion. The multiphase-flow frames of the Spurin data provide a second, real-geometry test regime: there the changes are fluid menisci rather than solid growth.
The scientific payoff is to recover, for every throat, a closure time and a radius-versus-time curve, and to test whether these curves can tell the two poisoning mechanisms apart when they are tuned to deposit the same total mineral mass. Bulk measurements, which is essentially what the calcium and nitrogen assays in the dental trial provide, cannot make this distinction.
Baselines
The naive reference is filtered back-projection at the same $K_t$, applied independently to each frame. It has negligible cost, and the 3DGR-CT comparisons show that it produces streaks and blurred voids under sparse views. Any method that fails to beat it on throat-radius error is useless.
The strongest existing method is 3DGR-CT fitted per frame. We run it in two forms. The first follows its paper, initialized from filtered back-projection and fitted to convergence. This form gives the quality ceiling at unbounded cost. The second, which is the most dangerous baseline, is warm-started from the previous frame's converged Gaussians, keeps its clone, split, and prune schedule, and receives exactly the wall-clock budget our method uses per frame, plus budgets 3× and 10× larger to trace a quality-versus-time curve. It uses the same projector, the same projections, and the same number of Gaussians, and its learning rates and density-control thresholds are tuned on the validation sequences with the same tuning budget we give our own hyperparameters. The only reported number for this family is from the medical benchmarks: 3DGR-CT "achieves PSNRs in the high 40s" at 80 projection views (3DGR-CT). We found no result on Lacuna for any Gaussian or learned reconstructor on time-resolved porous-rock tomography. We will therefore first reproduce 3DGR-CT on one medical volume from its benchmarks, to confirm our implementation is faithful, before trusting the rock comparisons.
We also include a learned voxel baseline, filtered back-projection followed by a 3D U-Net refiner, trained on the same simulated sequences and given the previous frame as an extra input channel. The ablations isolate the key ingredients. The first drops temporal carry-over, so each frame starts from filtered back-projection and the result is a per-frame ReSplat for CT. The second drops the birth head, so only offsets are applied to existing Gaussians. The third drops the residual input, so the network predicts the next state from the previous state alone, a pure dynamics prior. Beating the warm-started 3DGR-CT means lower throat-radius error and lower closure-time error at matched wall-clock, and quality within about 1 dB PSNR of the converged fit at a small fraction of its cost.
Why not just…?
Why not just warm-start 3DGR-CT from the previous frame and run a few iterations? This is the real threat. Consecutive frames are nearly identical, so warm starting may already remove most of the per-scan overhead that motivates this work. Our bet is that it fails in two specific places. The first is newly nucleated solid, where density control must grow Gaussians by cloning from possibly distant neighbors. The second is very low $K_t$, where a few gradient steps overfit the handful of views. The matched-budget quality-versus-time curve settles the question directly. If warm starting matches us in closure-time error at equal cost, the learned update adds nothing. In that case we would change the proposal: keep the optimizer and learn only the birth proposals as an initializer for it.
Why not just refine filtered back-projection with a 3D U-Net? This approach is fast and standard in sparse-view CT, and on voxel PSNR it may come close. Where we expect it to fall short is thin throats near voxel scale, and generalization when $K_t$ or the angle schedule changes at test time. A convolutional refiner is tied to the artifact pattern of one acquisition geometry. Our update, by contrast, re-reads the actual projection mismatch at every step, and ReSplat's 5 dB gain at unseen resolutions is the analogous evidence for residual feedback. The benchmark includes held-out view counts and held-out angle schedules specifically to test this.
Why not just take more projections, or rotate the sample more slowly? More projections per frame means longer exposure per frame, which blurs exactly the fast events of interest, meniscus snaps and closure moments. It also raises beam dose, which can itself perturb precipitation chemistry. The 2-second frames in the Spurin data are already a compromise. We report quality against $K_t$ so that experimenters can pick their own trade-off, rather than assuming sparse acquisition is free.
Why not reconstruct once and simulate the rest with Lattice Boltzmann or a GNN? Forward solvers such as the Lattice Boltzmann method used in the shale imbibition-flowback study, or LSE-MGN, need prescribed chemistry and a fixed geometry. Crystal growth poisoning is precisely the unknown: which sites are blocked is what we want to measure. Our reconstructions could instead supply time-varying geometry to those solvers, which addresses LSE-MGN's static-mesh assumption rather than competing with it.
Proposed Experiments
The first experiment is the matched-budget comparison on simulated precipitation sequences over held-out rock volumes. We report volume PSNR and SSIM, per-throat radius error on throats extracted with the same pore-network algorithm, closure-time error, error in the Euler characteristic (a count of connectivity) over time, and milliseconds per frame. The second tests transfer to real geometry on the Spurin multiphase frames. We subsample their fully sampled projections to $K_t$, treat the full reconstruction as ground truth, and score fluid-interface position against it.
The decisive check is mechanism identification. For pairs of sequences generated with uniform versus site-blocking poisoning at equal total mineral mass, a simple classifier is trained on our reconstructed closure-time curves and a second on the fully sampled ground truth. We ask whether the first comes within a few points of the second's accuracy at $K_t$ of 20 or below, while filtered back-projection and budget-matched warm-started 3DGR-CT fall clearly short. A downstream check converts the reconstructed throat radii into capillary entry pressures through $P_c = 2\gamma\cos\theta / r$, where $\gamma$ is interfacial tension and $\theta$ is contact angle, and counts how many throats flip between open and closed relative to ground truth. This is the error that matters for trapping predictions. If our method wins on PSNR but not on these throat-level measures, the claim fails as stated.
References
- Inhibitory effects of 1-hydroxyethane-1,1-bisphosphonate (HEBP) and Hexametaphosphate on the early calculus formation.
- Learning Pore-scale Multi-phase Flow from Experimental Data with Graph Neural Network
- 3DGR-CT: Sparse-view CT reconstruction with a 3D Gaussian representation
- ReSplat: Learning Recurrent Gaussian Splats
- GS-LRM: Large Reconstruction Model for 3D Gaussian Splatting
- DepthSplat: Connecting Gaussian Splatting and Depth
- The Influence Mechanism of Pore Structure of Tectonically Deformed Coal on the Adsorption and Desorption Hysteresis
- Reconstruction of 3D Porous Media From 2D Slices
- Lattice Boltzmann modeling of the coupled imbibition-flowback behavior in a 3D shale pore structure under reservoir condition