PaperlayerView exact revision on arXiv

Source-linked AI summary

Data-efficient crack quantification in lithium-ion cathodes using foundation model transfer

Thorsten Tegetmeyer-Kleine, Thomas Schmitt, Phillip Aquino, Christiane Rahe, Dirk Uwe Sauer, Weihan Li

arXiv:2608.27162v1cond-mat.mtrl-scics.CVcs.LG

TL;DR

Quantifying particle cracking in lithium-ion cathodes is limited by destructive microscopy and labor-intensive expert annotation. This paper uses frozen foundation-model features with lightweight decoding and iterative annotation to measure crack morphometrics, finding substantially more late intergranular cracking and tortuous networks after cycling than after initial or calendar-aged conditions.

  • Problem

    Quantifying particle cracking in lithium-ion cathodes is bottlenecked by destructive cross-sectional microscopy and labor-intensive expert annotation.

  • Method

    A frozen self-supervised vision-transformer encoder, lightweight decoder, and iterative model-assisted annotation infer per-particle morphometrics across three crack classes.

  • Results

    4.58% late intergranular crack area fraction occurred in the cycled-aged sample versus 0.52% initially and 0.49% after calendar aging, with more tortuous networks after cycling.

  • Takeaways & Limitations

    A single destructive cross-section can provide population-scale per-particle crack statistics for battery lifetime research and aging assessment.

  • Takeaways & Limitations

    Because only one SEM cross-section represented each aging condition, the statistics describe within-cross-section heterogeneity rather than cell- or lot-level population inference.

Abstract

from arXiv · show

Battery lifetime is central to sustainable electrification, yet the particle cracking that drives lithium-ion cathode aging is hard to measure: quantitative microscopy of this degradation is bottlenecked by annotation, because each destructive electron-microscopy cross-section spans hundreds of megapixels and pixel-level expert labelling requires hours per image. We show that a frozen self-supervised vision-transformer encoder, combined with a lightweight trainable decoder and iterative model-assisted annotation, turns this sparse labelling budget into population-scale degradation measurements. Applied to three 120-megapixel NMC cathode cross-sections representing initial, cycled-aged and calendar-aged states, the framework distinguishes intragranular cracks from early- and late-stage intergranular cracks and yields per-particle distributions of crack width, tortuosity and area fraction. Late intergranular crack coverage reaches 4.6% in the cycled sample versus 0.5% in the initial and calendar-aged samples, forming more tortuous, higher-coverage networks, consistent with degradation from repeated electrochemical cycling rather than elevated-temperature storage alone. A single destructive image yields the population-level statistics needed for lifetime-extending design, aging assessment and second-life decisions.

1. Introduction · 2. Results · Multi-class segmentation performance

The study addresses destructive, annotation-intensive crack quantification in NMC cathodes with a sparse-label framework using a frozen self-supervised vision-transformer encoder and lightweight decoder. Applied to three aging conditions, the model resolves three crack classes with population-scale morphometric potential despite contrast variation.

  • 1. Introduction: Destructive high-resolution microscopy and labour-intensive expert annotation limit quantitative tracking of cathode cracking and degradation evolution.FIB-SEM and nano-CT resolve fine details but are destructive, whereas laboratory micro-CT is repeatable but generally lacks resolution for small cracks.
  • 1. Introduction: Three crack classes distinguish intragranular, early intergranular, and late intergranular morphologies associated with different structural locations and degradation mechanisms.Intragranular cracks traverse individual grains, while intergranular cracks propagate along grain boundaries; the framework targets semantically distinct early- and late-stage intergranular cracks.
  • 1. Introduction: The framework freezes a self-supervised vision-transformer encoder and trains only a lightweight decoder, concentrating scarce annotation on task-specific adaptation.An iterative model-assisted workflow expands 79 hand-labelled SEM tiles into a larger expert-corrected dataset.
  • 2. Results: Three SEM cross-sections represented initial, cycled-aged, and calendar-aged NMC cathodes, each spanning approximately 17k×7k pixels at 22.3 nm/px.The samples showed condition-dependent differences in SEM contrast, brightness, and crack density, with pixel-level annotations defining the three crack classes.
  • Multi-class segmentation performance: 79 expert-labelled seed tiles produced a mean Dice score of 0.49 before model-assisted correction expanded the dataset to 482 expert-corrected tiles.The expanded set contained 79 fully manual seed tiles and 403 model-assisted tiles distributed across the three cross-sections.
  • Multi-class segmentation performance: 0.580 mean Dice was achieved across the three classes, with intragranular 0.575, early intergranular 0.513, and late intergranular 0.653.Mean precision was 0.729 and mean recall was 0.485 on the expert-corrected tile set (n=482; τ=0.1), while outputs resolved all classes across the three conditions.
  • Multi-class segmentation performance: 0.587 mean clDice and boundary F1 increasing from 0.446 at d = 1 px tolerance to 0.600 at d = 2 px indicate that narrow-crack errors are partly boundary-shift effects.Many cracks are only 2–5 pixels wide, making pixel-exact Dice conservative for morphology-preserving predictions.
  • Multi-class segmentation performance: SAM2 achieved binary Dice scores of 0.019 without prompts and 0.025 with Otsu-derived point prompts, and cannot directly resolve the required crack morphologies.The evaluation used the same 482-tile reference set with a binary crack-versus-background target.

Per-particle morphometric quantification

Per-particle distributions of crack width, tortuosity, and area fraction distinguish degradation morphologies beyond pixel-overlap metrics. Cycling particularly increases and complicates late-stage intergranular cracking, while the measurements describe within-cross-section heterogeneity rather than population-level inference.

  • Morphometric measurements: Per-particle width, tortuosity, and area-fraction distributions extracted from class maps resolved degradation morphologies that pixel-overlap metrics alone could not capture.The framework therefore provided morphology-sensitive measurements at the particle level.
  • Late intergranular cracking: 4.58% late intergranular area fraction in the cycled-aged sample exceeded 0.52% initially and 0.49% after calendar aging, an approximately ninefold increase over initial.This pattern is consistent with progressive grain-boundary degradation during repeated electrochemical cycling.
  • Late intergranular cracking: Late intergranular cracks in the cycled-aged sample were markedly more tortuous, indicating more interconnected grain-boundary crack networks than in the initial and calendar-aged samples.Mean late-crack width increased modestly from 159 ± 22 nm initially to 167 ± 39 nm after cycling, without statistical significance.
  • Condition-level distributions: Combined-class distributions retained the cycled-aged shift toward larger widths and tortuosities, including a broader late-intergranular width tail beyond 200 nm.These condition-level trends persisted across threshold reruns.
  • Scope and sampling: ROI-level statistics characterized within-cross-section morphometric heterogeneity rather than cell- or lot-level population inference because each condition used a single cross-section.Intergranular ROIs included 97 initial, 69 cycled-aged, and 49 calendar-aged particle-scale regions; intragranular ROIs included 446, 610, and 623 connected components, respectively.

3. Discussion

The discussion links cycling to enriched late intergranular cracking and greater crack coverage and tortuosity, while presenting sparse-label morphometric inference as a population-scale measurement framework. It also emphasizes robustness, broader applicability, and limitations from the three-cross-section case study.

  • Degradation interpretation: Cycling strongly enriches late intergranular cracking in NMC, unlike elevated-temperature calendar aging, with greater crack area fraction and geometric tortuosity.The findings are consistent with progressive grain-boundary degradation during repeated electrochemical lithiation–delithiation and the development of more interconnected crack networks.
  • Degradation interpretation: Anisotropic lattice evolution during cycling is consistent with mechanically incompatible strain between differently oriented primary grains and intergranular crack development.The described evolution involves in-plane lattice contraction and c-axis expansion followed by collapse near the H2→H3 transition at high state of charge.
  • Methodological contribution: The framework extracts per-particle crack width, tortuosity and area-fraction distributions from sparse, imperfect SEM annotations using a frozen self-supervised encoder and lightweight decoder.Per-pixel class probabilities provide the intermediate representation for deriving these morphometrics while concentrating expert labeling on the scientific question.
  • Robustness: Robust morphometric trends persist across thresholds and descriptors, because width and tortuosity average over crack geometry rather than relying on exact pixel overlap.This supports degradation signatures beyond pixel-level segmentation metrics despite local annotation uncertainty.
  • Broader implications: Each destructive cross-section yields hundreds of per-particle morphometric measurements, enabling population-level crack characterization and potential application to other destructive-microscopy domains.The framework is presented as relevant to manufacturing quality assurance, aging diagnostics and quantification of single-crystal NMC designs.
  • Limitations: The case study evaluates one frozen vision-transformer encoder and lightweight decoder across only three SEM cross-sections, leaving systematic architecture and training-seed benchmarking for future work.The authors frame the sparse-label morphometric-inference framework, rather than optimization of a specific segmentation architecture, as the primary focus.

4. Conclusions · 5. Methods · Multi-class segmentation of crack types

The paper introduces a sparse-label framework that uses frozen self-supervised vision-transformer features and a trainable decoder to infer per-particle crack morphometrics. Crack typing is formulated as three-class semantic segmentation, supported by iterative expert-corrected annotation and calibrated operating-point evaluation.

  • 4. Conclusions: The framework converts few expert annotations into per-particle crack morphometrics while concentrating annotation effort on quantification rather than representation learning.It uses frozen self-supervised vision-transformer features rather than learning the representation from scratch.
  • 4. Conclusions: 79 hand-labelled tiles grew into a 482-tile expert-corrected reference set through iterative model-assisted annotation across three NMC cathode cross-sections.The framework resolved intragranular as well as early crack categories in the supplied conclusion passage.
  • Multi-class segmentation of crack types: Pixel-wise crack typing is formulated as three-class semantic segmentation with intragranular, early intergranular, and late intergranular labels.The intragranular class reflects defect-induced crack initiation within primary particles.
  • Multi-class segmentation of crack types: 85.7 M frozen parameters and 768-dimensional patch embeddings come from the DINOv3 ViT-B/16 encoder, while the DeepLabV3+ decoder has 17.8 M trainable parameters.Each greyscale 1024 × 1024 px SEM tile is replicated to three channels and produces a 64×64×768 feature map from 16 × 16 patches.
  • Multi-class segmentation of crack types: τ=0.1 was selected on the validation precision–recall curve, after which binary masks were obtained by thresholding class probabilities.Training minimised a weighted sum of binary cross-entropy, mean-squared-error, and Dice losses.
  • Multi-class segmentation of crack types: 80%/20% per-image partitioning yielded 96 held-out validation tiles from the 482-tile pool, comprising 79 manual seed tiles and 403 model-assisted, expert-corrected tiles.The 403 corrected tiles were distributed as 160 cycled-aged, 81 calendar-aged, and 162 initial tiles; metrics on all 482 tiles served operating-point calibration.

Baseline evaluation using SAM2 · Inference and overlay mask generation · Segmentation mask analysis pipeline

The section evaluates SAM2 against the DINOv3 framework on binary crack segmentation, then describes full-resolution inference, overlay generation, and morphometric extraction from predicted masks. The pipeline converts likelihood maps into calibrated particle-level crack geometry and width measurements.

  • Baseline evaluation using SAM2: SAM2 was evaluated using automatic dense-grid mask generation and prompted segmentation with Otsu-thresholded dark-region point prompts.The comparison used facebook/sam2.1-hiera-large in two configurations.
  • Baseline evaluation using SAM2: Because SAM2 is class-agnostic, the three reference classes were collapsed into a binary crack-versus-background target for comparison.This binary evaluation allowed comparison with the DINOv3 framework.
  • Baseline evaluation using SAM2: Dice = 0.77 was achieved by the DINOv3 framework against SAM2 values reported in Table A.1 on the binary metric.The reported metric is for crack-versus-background segmentation rather than the original three-class target.
  • Inference and overlay mask generation: Full-resolution inference on ∼17,000×7,000 px images used overlapping sliding windows with per-class prediction averaging across overlapping tiles.This strategy was applied to the initial, calendar-aged, and cycled-aged probes.
  • Inference and overlay mask generation: Final per-class binary masks were thresholded and rendered as alpha-blended overlays at native resolution for side-by-side probe comparison.The overlays supported visual comparison across the three aging states.
  • Segmentation mask analysis pipeline: The morphometric pipeline cleans each likelihood map into a binary crack mask, extracts particle-level ROIs, and derives skeleton-based width and geometry distributions.ROIs represent individual cathode secondary particles or small clusters.
  • Segmentation mask analysis pipeline: Measurements are computed in pixel units and converted to physical units through SEM calibration after thresholding m(x) at τ into B(x) = 1{m(x) ≥τ}.The thresholded binary field forms the basis for subsequent crack measurements.
  • Segmentation mask analysis pipeline: Skeletonisation reduces each ROI’s binary crack mask to one-pixel centrelines, while distance-transform sampling estimates crack width from maximal inscribed disks.Non-finite or nonpositive samples are excluded from summary statistics but retained as zeros in visualisation maps.

Geometric descriptors and width distributions

The section defines geometric descriptors for crack morphology, including topology, orientation, anisotropy, tortuosity and transverse roughness. It also characterizes crack-width distributions using descriptive statistics, histograms, ECDFs and quartiles for cross-ROI comparison.

  • Geometric descriptors: Topology classifies crack pixels by 3 × 3 neighbour degree into endpoints, branch points, and isolated pixels.Endpoints have degree 1, branch points degree ≥3, and isolated pixels degree 0.
  • Geometric descriptors: Orientation and anisotropy are derived from covariance eigenpairs of centered skeleton coordinates.Orientation is reported over [0°, 180°).
  • Geometric descriptors: Geometric tortuosity is the centreline-length-to-projected-span ratio along the principal direction, distinct from effective pore-network transport tortuosity.Because length uses a raw skeleton-pixel count, near-linear cracks can produce tortuosity marginally below unity without indicating measurement error.
  • Width distributions: Crack widths are summarized by minimum, maximum, mean, median, and standard deviation, with histograms and ECDFs visualizing their distributions.The ECDF gives the fraction of centreline length with width ≤t.
  • Width distributions: Quartiles P25, P50, and P75 are marked to support cross-ROI comparison and aggregated ECDF analysis.Aggregated ECDFs are presented across multiple panels of Fig. 5.

Statistical analysis · Reporting and assumptions · CRediT authorship contribution statement

The analysis compares per-ROI crack morphometrics across aging conditions with non-parametric tests and multiple-comparison correction, while reporting calibrated pixel-based measurements and threshold sensitivity. Contributions span conceptualization, methodology, analysis, data curation, supervision, funding, and manuscript review, with all authors approving the final manuscript.

  • Statistical analysis: Per-ROI area fraction, mean width, and geometric tortuosity distributions were compared between aging conditions.
  • Statistical analysis: Two-sided Mann–Whitney U tests assessed differences between aging conditions.
  • Statistical analysis: Cliff’s δ provided a non-parametric effect-size estimate, with Benjamini–Hochberg correction controlling family-wise multiple comparisons.
  • Reporting and assumptions: Quantitative measurements came from thresholded crack masks, with pixel values converted to physical units using SEM calibration.
  • Reporting and assumptions: Width statistics depend on τ, and sensitivity to this parameter was reported in Fig. A.2.
  • CRediT authorship contribution statement: Thorsten Tegetmeyer-Kleine led conceptualization, methodology, software, validation, formal analysis, data curation, visualization, and original-draft writing.
  • CRediT authorship contribution statement: All authors reviewed and approved the final manuscript, while other contributors provided conceptualization, investigation, resources, supervision, funding acquisition, and manuscript review.

Appendix A. Extended data … S1. Training configuration

The appendix validates threshold selection, compares segmentation baselines, and tests morphometric sensitivity, while supplementary methods specify sample preparation, SEM imaging, and frozen-encoder training configuration. Full-resolution inference uses overlapping windows, stitching, optional intensity gating, cleanup, and particle-level ROI extraction.

  • Appendix A. Extended data: A low operating threshold reflects extreme foreground–background imbalance, with crack pixels comprising <1% of the image.Per-class precision–recall curves and Dice-versus-threshold plots support threshold selection.
  • Appendix A. Extended data: Zero-shot SAM2 configurations were evaluated against fine-tuned DINOv3 on the same 482-tile reference set using binary Dice and IoU.Reference labels were collapsed from three classes to crack versus background, retaining the higher filtered or unfiltered SAM2 value per tile; confidence intervals used 1000 bootstrap resamples.
  • Appendix A. Extended data: Area fraction decreases monotonically as threshold τ increases, whereas mean crack width and tortuosity remain more stable across τ ∈ {0.05, 0.1, 0.2, 0.3, 0.5}.The appendix attributes this contrast to suppression of low-confidence pixels and skeleton-based averaging over retained crack pixels.
  • Appendix A. Extended data: Morphometric statistics are reported by sample and crack class as mean ± std across ROIs at a pixel size of 22.3 nm.These statistics support the per-condition morphometric comparisons in the main article.
  • Supplementary Notes: Three NMC cathode conditions were prepared from commercial Ni-rich NCM pouch cells: formation-only initial, K100 after 100 charge–discharge cycles, and K187 stored at 45°C.One cell per condition was prepared for post-mortem cross-sectioning.
  • Cross-section SEM imaging: SEM cross-sections were acquired at 10 kV using a ZEISS Supra 55 BSE detector, with tiles assembled without line or frame averaging at approximately 22.3 nm/px.The supplied passage also reports per-sample pixel sizes of 22.29 nm/px for K100, 22.38 nm/px for K187, and 22.28 nm/px for the initial sample.
  • S1. Training configuration: Only the decoder and output head were trained while the encoder remained fixed, using AdamW with η=10^-4, β1=0.9, β2=0.98, ε=10^-9, and λ=10^-3.Training used a cosine-cyclical schedule with 4 warm-restart cycles and a minimum learning rate of 5×10^-5.
  • S1. Training configuration: Training ran for up to 500 epochs with patience 20 early stopping, batch size 4 at 1024×1024 px in FP32, seed 42, and checkpoint selection at epoch 465.The objective combined per-channel weighted BCE, MSE, and Dice losses with equal weights λBCE = λMSE = λDice = 1.

S3. Segmentation operating point and threshold selection

All quantitative results use a decision threshold of τ=0.1, selected from the validation precision–recall curve to maximize class-averaged Dice for highly imbalanced thin-crack segmentation. Crack pixels comprise less than 1% of tile area, and higher thresholds reduce recall faster than false positives.

  • Threshold selection: τ=0.1 is the decision threshold used for all quantitative results in the main article.The operating point was selected on the validation precision–recall curve.
  • Threshold selection: Less than 1% of each tile area consists of crack pixels, creating highly imbalanced foreground fractions.This imbalance is characteristic of thin crack structures in SEM cross-sections.
  • Threshold selection: At higher thresholds, recall collapses faster than the false-positive rate is reduced.The threshold was chosen to maximize class-averaged Dice under this precision–recall trade-off.

S4. Supplementary morphometric definitions

The pipeline uses standard morphometric conventions for skeleton topology and crack-width distributions. Skeleton pixels are classified by local eight-connected neighbourhood degree, while width ECDFs are summarized with robust quartiles.

  • Skeleton topology: Skeleton pixels are classified on an eight-connected 3×3 window by neighbourhood degree: degree 1 endpoints, degree ≥3 branch points, and degree 0 isolated pixels excluded from path-length statistics.This is the standard medial-axis classification used throughout connected-component morphometry.
  • Width statistics: Width statistics use the empirical cumulative distribution function, defined as the fraction of centreline length with width ≤t.The width values are derived from the Euclidean distance transform.
  • Width statistics: Width ECDFs are summarized by P25, P50, and P75 rather than mean and standard deviation because distributions are typically heavy-tailed and right-skewed by junction pixels.These quartiles are used as robust descriptors for cross-ROI comparisons.

S5. Spatial overlap quantification · S6. SAM2 baseline details · S7. Statistical-analysis bookkeeping

The supplementary analyses define how spatial overlap, SAM2 baseline outputs, and multiple-comparison statistics were handled. Validation metrics reflect overlapping within-image tiles, SAM2 was evaluated class-agnostically, and the headline cycling-versus-baseline claim survived FDR adjustment at q < 0.001 in both test families.

  • S5. Spatial overlap quantification: Each validation tile was assigned an overlap fraction computed over training tiles from the same SEM cross-section.The tiling used 50%-overlap sliding windows.
  • S5. Spatial overlap quantification: 50% adjacent-tile area sharing and the random 80/20 partition made the median overlap fraction across 96 validation tiles ≈0.50 by construction.The stated probability that all four cardinal neighbours are validation tiles is 0.24 ≈0.002.
  • S5. Spatial overlap quantification: Image-disjoint splits were omitted because only three SEM cross-sections were available, one per aging condition, making leave-one-condition-out validation infeasible without losing condition coverage.A rotating image-disjoint sensitivity analysis was left for future work pending additional independent cross-sections per condition.
  • S6. SAM2 baseline details: SAM2 was evaluated in automatic mode with a 64 × 64 dense point grid and in prompted mode with Otsu-derived centroid prompts, allowing up to 128 prompts per tile.Prompted candidates were area-bounded and selected by lowest mean intensity, with candidate area ≤0.25 of the tile as stated in the passage.
  • S6. SAM2 baseline details: SAM2 outputs were compared against a binary crack-versus-background target by collapsing the three-class reference, while heuristic width-based class assignment was rejected.The rejection avoided conflating SAM2 segmentation quality with the supervised classifier.
  • S7. Statistical-analysis bookkeeping: 27 comparisons covered 3 ROI subsets, 3 morphometric metrics, and 3 condition pairs, with Benjamini–Hochberg-adjusted q-values reported within this family.ROI counts were 194/138/98 for the intergranular early subset as stated in the passage.
  • S7. Statistical-analysis bookkeeping: The separate 9-test intergranular-late family also used Benjamini–Hochberg adjustment, and the headline cycling-versus-baseline claim survived FDR adjustment at q < 0.001 in both families.The intergranular-late ROI counts were 97/69/49.

S8. Software environment and reproducibility · Supplementary Figures

The supplementary materials specify the training environment, model configuration, augmentation and inference settings, and automated ROI-extraction criteria. Supplementary figures document skeleton-based morphometrics, boundary-F1 behavior, method diagnostics, ROI detection, and segmentation performance across crack densities.

  • S8. Software environment and reproducibility: All models ran on a single NVIDIA A100 80 GB PCIe GPU under Ubuntu 22.04.5 LTS with Python 3.10.13 and the listed deep-learning and image-analysis packages.The environment used PyTorch 2.4.0 with CUDA 12.1 and cuDNN 9.1.0, torchvision 0.19.0, PyTorch-Lightning 2.0.8, timm 0.9.6, scikit-image 0.25.2, SciPy 1.15.3, NumPy 1.26.4, and OpenCV.
  • S8. Software environment and reproducibility: The frozen DINOv3 ViT-B/16 encoder fed a DeepLabV3+ decoder with ASPP dilation rates 6/12/18 and low-level stride 2.The configuration had 17.8 M trainable parameters and 85.7 M frozen parameters.
  • S8. Software environment and reproducibility: Training used online augmentation only on the training split, while validation data remained unaltered.The Dice loss used smoothing constant ε=1, and augmentation operations were independently sampled at their specified probabilities, including random affine transforms at probability 0.5 with rotation up to ±45◦.
  • S8. Software environment and reproducibility: Inference binarised per-class probability maps at τ=0.1 and extracted particle ROIs using blurred backscatter thresholding, morphological closing, fragment merging, and minimum component criteria.The pipeline used a 5 × 5 Gaussian pre-blur, a 5×5 elliptical kernel, a 50 px merge radius, particle area ≥10,000 px, crack area ≥1000 px, and skeleton length ≥100 px.
  • Supplementary Figures: Skeleton-based extraction converted segmented crack masks within each ROI into medial-axis skeletons with branch points and endpoints for morphometric analysis.Supplementary Figure S1 illustrates the procedure for an intergranular late ROI, where the crack mask covered 16.1% of pixels, or 17 666 px.
  • Supplementary Figures: Median boundary-F1 was 0.45–0.46 per class at 1-pixel tolerance and 0.58–0.62 at 2-pixel tolerance, indicating that 1–2-pixel boundary misalignment drives moderate pixel-exact Dice.The comparison covers intragranular, early intergranular, and late intergranular crack classes.
  • Supplementary Figures: Supplementary diagnostics compare expert references with SAM2-auto, SAM2-prompted, and fine-tuned DINOv3 predictions using binary Dice and error maps.The diagnostic is shown for tile 100_0000_00038, the Fig. 1 sample tile, with cyan true positives, magenta false positives, amber false negatives, and black true negatives.
  • Supplementary Figures: Across 482 expert-corrected tiles, Dice and IoU were plotted against foreground pixel fraction for intragranular, early intergranular, and late intergranular cracks, with Pearson r = 0.21–0.32.The supplementary figures also show the automated ROI-detection workflow and its thresholding, post-merge, and rejected-component stages for the cycled-aged image.
Loading 2608.27162v1…