Recovering Sub-Voxel Metal Deposits in Operando Tomography by Referencing the Pristine Scan

When a lithium or zinc battery charges, metal plates onto the electrode. If it plates unevenly it grows needle-like dendrites that short the cell. Materials chemists try to prevent this with "templated" surfaces: chemical groups or seed particles that metal ions prefer to attach to, so the first layer of metal forms where the engineer wants it. The obvious way to check whether a template works is to film it. Synchrotron X-ray tomography produces a 3D scan of a working cell at each of many time points, which amounts to a volumetric video. The best existing analysis of such a video, batteryNET, segments each 2D slice on its own. It cannot account for more than 70% of the lithium that Faraday's law says was deposited (Huang et al.). That missing lithium is the part that matters most for templating. It sits in nucleation layers thinner than the 1.33 µm voxel. In any single scan it shows up only as a faint change in gray level.

This note proposes to treat the first, pristine scan of the cell as a per-voxel reference. The segmenter then predicts, for each voxel at each later time step, the fraction of that voxel occupied by new metal, not a hard class label. The faint signal in one noisy scan is below the noise. It becomes measurable when it is registered against the pristine state, persists across consecutive time steps, and is pooled over connected regions. Charge passed through the cell, which the potentiostat records for free, gives an upper bound on how much metal can exist at each time. The scientific payoff is a quantitative test of the open question in templated nucleation: does new growth keep originating at template sites over cycling, or do the templates get buried and stop working?

Core Claim

Hard-label segmentation of each time point cannot see sub-resolution nucleation. A model that is anchored to the pristine scan, temporally coupled, and outputs voxel fractions can recover a substantial share of Faradaically deposited metal that current pipelines miss. The proposal does not invent mass to do this: the Faraday bound only caps the total and never rewards adding deposit. Recovering that mass makes it possible to attribute each growth event to the pristine interface site it started from. This turns the "structural memory" that Huang et al. observed qualitatively into a measured quantity. Their observation was that dendrites grow from defects already present in the pristine cell. The measured quantity is the fraction of new deposit originating at template sites, as a function of time.

Position Relative to Prior Work

Three papers come closest, and each is missing one ingredient.

batteryNET (Huang et al.) supplies the data regime and motivates the gap. It segments 25 operando scans into five hard classes with a 2D iterative residual U-Net, reaching 98.8% accuracy and 95.5% IoU. It also documents both the structural-memory effect and the >70% capacity gap. We keep its classes and labeled slices. We change the output from a hard label to a sub-voxel fraction and replace independent slices with a pristine-referenced time series.

3D dendrite segmentation of alloy solidification (James et al.) supplies a method and a constraint. Using 3D context let a 3D U-Net ignore ring-streak artifacts that confuse 2D networks, reaching 99.84% pixel accuracy and 0.58-pixel boundary error. A context of only four z-slices was optimal, which warns that longer context is not automatically better. Their volumes are static, so time plays no role. We borrow the 3D backbone, the noisy-then-clean fine-tuning recipe and the hyperparameter search. We run the analogous temporal-window ablation, expecting a similar sweet spot.

Residual ConvLSTM (Lee et al.) supplies simulated ground truth and a negative result. It predicts dendrite growth on 90,000 phase-field sequences (simulations of the evolving metal–electrolyte boundary) and learns per-step changes. It is a forward simulator, not a segmenter, and it degrades at 0.5 V where branching becomes stochastic. We reuse the idea of modeling change between steps. We use its simulation setup as a source of known sub-voxel truth, not as a model to beat.

Volumetric video segmentation contributes a principle rather than a drop-in method. STEm-Seg (Athar et al.) clusters voxels of a space-time volume into Gaussian instance embeddings, avoiding a separate detect-then-track stage. Flow-free affinity methods (Keuper et al.; Zou et al.) were designed because optical flow breaks under noise. Both assume objects persist and move. Electrodeposition instead creates new material at fixed locations. The right temporal prior is therefore "accumulation relative to a reference," not "identity carried along a trajectory," and this is the main departure from that literature. We reuse the Gaussian-centroid idea only for the attribution step: each growth instance gets a space-time center whose earliest appearance is linked to a pristine-interface site.

The templating papers define the hypotheses to be measured. These include carbonyl zincophilic sites (Jiang et al.), sodiophilic oxygen-doped carbon foam (Cui et al.), Sb/Sb-Zn alloy interfaces (Zheng et al.) and Bi seed nanoparticles reduced from the electrolyte (Cui et al.). All infer template function from overpotentials, cycling and post-mortem images. None tracks, in a living cell, whether growth keeps starting at the template once it is covered. That missing measurement is the open problem the method targets.

Research Program

Output representation. The model reads a registered pair or short window: the pristine volume plus the volumes at times $t-k$ through $t$. It outputs the five batteryNET classes and, for the metal classes, a fraction $\phi_v(t)\in[0,1]$ of voxel $v$ occupied by metal.

Registration. Registration is the first engineering problem. The electrode gap swells by roughly 39.3 µm over a charge (Huang et al.), so each scan is deformably aligned to the pristine scan. The alignment uses features away from the active interface, such as the current collector and the bulk solid electrolyte. This keeps the alignment from erasing the very changes being measured.

Training losses. Training combines four terms:

  • cross-entropy on the hand-labeled slices;
  • a partial-volume supervision term from simulation, described below;
  • a temporal term penalizing metal fraction that appears and disappears between adjacent steps during a monotonic charge segment;
  • the Faraday bound.

The Faraday bound. Let $Q(t)$ be the cumulative charge passed, $M$ the molar mass, $z$ the ion charge number, $F$ Faraday's constant, $\rho$ the metal density and $\Delta$ the voxel edge length. The bound is

$$\sum_v \phi_v(t)\,\Delta^3 \;\le\; \frac{Q(t)\,M}{z F \rho},$$

applied as a one-sided hinge penalty. It is one-sided because part of the charge goes into interphase chemistry rather than metal. A two-sided constraint would therefore force the model to hallucinate deposit. Any gap that remains after segmentation is an upper estimate of charge lost to side reactions, which is itself a useful number.

Simulated ground truth. We generate it by rerunning the phase-field setup of Lee et al.. Their idealized planar electrode is modified to carry seeded sites of controlled density and binding strength, mimicking templated substrates. Each fine-grid field is rendered through a tomography forward model:

  • blur to a 1.33 µm point spread;
  • Poisson noise;
  • ring artifacts.

Grid spacing is chosen so that early nuclei are smaller than one output voxel. This yields ground truth for exactly the regime real data cannot label: fractional occupancy, true deposited mass, and which seed each growth event came from. The rerun is required work, since the published dataset has no surface templates.

Attribution. Each connected growth region is assigned a space-time centroid in the manner of Athar et al. and traced back to its earliest time of detectable fractional deposit. It is then linked to the nearest feature on the pristine interface. The quantity reported per time step is the fraction of newly deposited metal attributable to pre-identified template or defect sites. Template burial would appear as this fraction decaying across cycles.

Real data. The real-data program has two stages. The first applies the model to the batteryNET cycle, conditional on obtaining those scans from the authors. The second proposes new operando acquisitions of cells with and without a templated host, so that the attribution curve can be compared between conditions.

Baselines

Naive reference: batteryNET as published. This is the 2D slice-wise iterative residual U-Net (Huang et al.), retrained on the same 42-slice-equivalent label budget. Its reported numbers are 98.8% accuracy, 95.5% IoU, and more than 70% of Faradaic capacity unobserved. Beating it means two things. First, the method must keep IoU on mature dendrites, pits and deposits within about a point of 95.5% on held-out labeled slices. Second, it must account for a substantially larger share of the deposited charge on the same cycle.

Strongest per-time-point learned competitor: 3D U-Net. This follows James et al.: applied independently to each scan, pre-trained on noisy labels and fine-tuned on the clean ones. It gets the same labeled data, the same GPU budget and the same number of hyperparameter trials under the same asynchronous successive-halving search they used. Their reported 99.84% accuracy and 0.58-pixel boundary error are on Al-Zn solidification, not batteries. No reported result was found on Lacuna for any method recovering sub-voxel deposit mass. On simulation, the comparison is mass-recovery error and nucleus detection recall for nuclei smaller than one voxel. On real data, it is the change in Faradaic closure and the temporal label-flicker rate.

Strongest non-learned competitor: register and subtract. The pipeline deformably registers each scan to the pristine one, subtracts, smooths with a Gaussian, and maps intensity difference to metal fraction using a calibration from the simulations. It gets the same registration as our model, so only the learned part differs. This baseline is the most dangerous, as discussed below.

Ablations isolating the key ingredients. The full model is compared against four variants: without the pristine input; with hard labels instead of fractions; without the temporal term; and without the Faraday hinge. A final variant removes the hinge and replaces it with a two-sided Faraday constraint. That variant should overfill the gap on simulations where true side-reaction loss is known. It demonstrates why the bound must be one-sided.

Why not just…?

Why not just run a 3D U-Net on each scan and lower its threshold? A single scan carries the partial-volume signal of a sub-voxel nucleus at roughly the noise level, so lowering the threshold adds false positives as fast as true deposits. James et al. show 3D context suppresses structured artifacts, but nothing in a single frame separates a faint new deposit from a faint pre-existing feature. The pristine reference is what does that. The simulation experiment settles this directly. If the per-scan 3D U-Net, with its threshold swept, matches our nucleus recall at equal precision below one voxel, the temporal reference is unnecessary.

Why not just subtract the registered pristine scan and threshold? This is a real threat. If registration is good and artifacts are stationary, a difference image may recover most of what our model does. A learned model should only help where non-stationary ring artifacts, swelling-induced residual misalignment, or overlapping classes such as redeposited lithium versus dead lithium corrupt the difference. We run this baseline with identical registration and report it as a first-class result. If it matches the learned model on simulated mass recovery and real Faradaic closure, the contribution becomes the measurement protocol itself: pristine referencing, fractional output, the one-sided charge bound and attribution. The network would be dropped.

Why not just apply a video segmenter like STEm-Seg, or optical-flow propagation, to the time series? STEm-Seg (Athar et al.) reports 34.6 AP on YouTube-VIS and 64.7 J&F on DAVIS 2019. It and flow-based temporal fields (Bao et al.) assume objects persist and move, and they output hard masks. Electrodeposition creates material in place, and the missing mass is sub-voxel. Smoothing hard labels over time reduces flicker but cannot add fractional mass. The hard-label ablation tests this. If temporal smoothing of hard labels closes as much of the Faradaic gap as fractional output, our representation claim fails.

Why not just image at higher resolution? Higher resolution would remove the need for sub-voxel inference, but it shrinks the field of view. Operando cells need the full interface to measure where growth originates. Where small-field higher-resolution scans of the same cell are obtainable, we use them as an independent check of predicted fractional deposits, not as a replacement.

Decisive Checks

The first decisive check is on simulation. On held-out phase-field sequences rendered through the forward model, the full model must recover total deposited mass within a stated tolerance of truth. Beyond that, it must beat both the per-scan 3D U-Net and the subtraction baseline on recall of nuclei smaller than one voxel, and on attribution accuracy to the correct seed. We will also report where all methods fail. The stochastic high-overpotential regime that broke the ConvLSTM of Lee et al. is the expected failure point for attribution.

The second check is on the simulated burial curve. The model must reproduce the known decay of template-originated growth across simulated cycles. We vary template density and binding strength to confirm it does not output a decay regardless of the input.

On real data, success has three parts:

  • no loss of mature-class IoU against batteryNET's 95.5%;
  • a larger accounted share of Faradaic charge, with the hinge never active, meaning the model never overshoots the charge bound;
  • lower label flicker in regions with no electrochemical activity.

The final scientific test requires the proposed acquisitions. The template-attribution fraction should differ between templated and untemplated cells in the direction the templating literature predicts. Whether it decays across cycles is the open question this work is built to answer.