Calendering Defects in Li-Ion Electrodes: A Detection Guide

This article accompanies our video on calendering defect detection. The video covers the visual signatures of each defect type. This piece goes deeper on process root causes and detection benchmarks.

Calendering is the roll-pressing step that compacts a coated electrode from roughly 45–60% porosity down to 25–35%, setting the final density, thickness, and mechanical integrity of the active layer before slitting. When it goes wrong, you get six recurring defect families: edge cracking, delamination, pinholes, thickness non-uniformity, wrinkling, and active material fracture. Each one degrades cell capacity, raises lithium plating risk, or scraps the electrode outright. Detection matters because most of these defects don't show up in cell test data until weeks later, after formation, when the electrode is already inside a sealed can. At Niobia AI, we've built our detection and root cause analysis platform specifically around this lag problem.

What is calendering and why does it affect electrode quality?

Calendering passes a coated foil, typically aluminum at 15–20 μm for cathodes or copper at 6–12 μm for anodes, between two heated rolls under line loads that can reach 5.9 kN/mm. The pressure compacts the porous active layer, increases volumetric energy density, lowers electronic resistance, and improves particle-to-particle contact. Production line speeds run 60–100 m/min, with roll temperatures ranging from 25 to 145 °C depending on chemistry and binder system^[1].

The physics is straightforward but unforgiving. Meyer, Bockholt, Haselrieder, and Kwade^[2] showed that compaction follows a Heckel-type exponential, where electrode density rises sharply at first then plateaus. For NMC111, going from 60% to 37% porosity requires roughly 60–200 MPa of apparent pressure. The catch is that "apparent" pressure is a fiction. Force chains through the granular bed mean some particle contacts see roughly 6× the average force, which is exactly why polycrystalline NMC811 secondary particles fracture at industrial pressures even though the bulk fracture stress sits in the 6–14 GPa range^[3].

Get it right and you get a denser, better-conducting electrode. Push too hard and you collapse the pore network, fracture cathode particles along grain boundaries, and seed lithium plating sites on the anode. The window between optimal and over-calendered is narrow, often just a few microns of thickness change.

The six main defect types from calendering

The reference taxonomy is Günther et al.^[4], which sorts calendering defects into geometric, mechanical, and microstructural classes. Here's what you actually see on the production line.

Edge cracking appears as fissures running parallel to the web direction at the coating-to-bare-foil boundary, or as transverse cracks across the active layer. The mechanism is differential plastic strain. The coated zone elongates by up to 5% during compaction while the uncoated foil edge does not, generating tensile shear at the interface. On NMC811 with high active-mass loading, edge cracking shows up first as a hairline whisker visible only under raking light. Left unchecked, it propagates into delamination during slitting.

Delamination is the active layer detaching from the current collector across a contiguous area, usually a few millimeters wide or larger. It's distinct from coating peel, which is a surface-level lift-off of a thin top layer often caused by binder migration during drying. Delamination typically traces back to inadequate adhesion combined with shear stress from over-compaction, and you find it via 90° peel tests, ultrasonic NDE, or by spotting blank foil under inline brightfield illumination.

Pinholes are small voids reaching the current collector. After calendering, they often present as high-gloss spots because the surrounding active material has been smoothed flat while the pinhole edge stays matte. Inline optical systems reliably resolve pinholes down to roughly 50–100 μm at production speed. Anything smaller needs beta-gauge thickness mapping or X-ray. Root causes are slurry air entrapment plus binder migration during drying that gets sealed in by the rolls.

Thickness non-uniformity is the bread-and-butter calendering defect. Schreiner et al.^[5] report a mean simulation-to-experiment thickness deviation of ±2.88 μm for NMC622 cathodes. Industry tolerance for high-end cells runs ±2 to ±5 μm across the full web width. The cause is upstream coating variation, agglomerates causing local roller deflection, or the simple fact that material at the web edge flows freely while material in the center does not.

Wrinkling is out-of-plane buckling of the foil, usually showing up as longitudinal corrugations near the uncoated edge. It's a tension-mismatch defect. The coated zone elongates under compaction, the uncoated foil does not, and if web tension isn't balanced you get buckling. Production webs typically run at 30–200 N/m tension, and the working window narrows considerably for thin copper foil at the 6 μm end of the range.

Active material fracture is the microstructural defect that doesn't announce itself visually. Polycrystalline NMC811 secondary particles crack along primary-crystallite grain boundaries at typical industrial pressures because of those force-chain stress concentrations. Particles larger than 10 μm are most vulnerable. Ngandjong et al.^[6] used DEM simulation plus nano X-ray CT to show fracture progressing through three regimes: rearrangement, partial cracking, complete crushing. Graphite anodes start breaking up around 400–800 MPa (4–8 t/cm²), which kills first-cycle Coulombic efficiency and doesn't show up in visual inspection at all.

What process parameters drive each defect type?

Four parameters do most of the damage when they drift: roll gap, roll temperature, line speed, and incoming coating uniformity.

Roll gap drives porosity, and porosity drives almost everything else. NMC cathodes target 25–35% post-calender porosity. LFP runs in the same band, often 25–35%, because LFP's poor electronic conductivity demands denser packing. Graphite anodes target around 30%, with 25–35% as the working range. Push below 25% on a graphite anode and you start choking ion transport. Hidalgo et al.^[7] characterized NMC622 across 30/35/40% porosity targets at 85, 120, and 145 °C, finding that lower porosity helps high-rate performance only down to a point. Below roughly 27%, you create what the literature calls a "lithium-deficient dead zone."

Roll temperature controls binder deformability. PVDF gets noticeably more compliant above 80 °C, which is why NMC cathodes calendered at 85–145 °C compact more uniformly than cold-calendered ones. A 10 °C variation across the roll face produces a thickness gradient that beta-gauge picks up but visual inspection misses entirely.

Line speed interacts with temperature and dwell time. Run faster and the active material sees less effective compaction at a given line load, which manufacturers compensate for by raising pressure, which then increases particle fracture risk. The drift pattern that precedes a defect cluster is usually thickness creeping by 1–2 μm over 30 minutes while everything else looks nominal.

Incoming coating variability is the silent killer. Schreiner et al.'s ±2.88 μm post-calender deviation is downstream of whatever your slot-die coater is doing upstream. If your coating thickness varies by ±5 μm before calendering, the calender amplifies that into a porosity gradient because pressure equalization is imperfect. This is the propagation effect that most quality systems miss, and it's one of the first things Niobia AI maps when onboarding a new line.

How are calendering defects currently detected?

The detection stack on most production lines has four layers, with very different resolutions and costs.

Manual visual inspection still anchors many lines, particularly during ramp. The data is not flattering. Human factors research puts inspector accuracy at roughly 80% under good conditions, with 20–30% miss rates after two hours of continuous observation^[8]. On a moving electrode web at 60–80 m/min, humans reliably catch pinholes and macro-defects above 100–200 μm. Smaller features and any subsurface microstructural change are invisible.

Laser profilometry does the heavy lifting on thickness and 3D surface defects. Keyence LJ-X8000 series sensors offer 2.5 μm lateral resolution, sub-micron Z repeatability, and profile rates up to 64 kHz. Micro-Epsilon scanCONTROL units run up to 4,000 Hz with 12 μm point spacing. They're excellent for thickness mapping, edge defect detection, and corrugation amplitude measurement. They're poor at flush pinholes and subsurface particle fracture.

Inline optical inspection uses line-scan cameras with combined brightfield and darkfield illumination. Schoo et al.^[9]describe an 8,192-pixel dual-channel system at 35 × 37.5 μm pixel resolution resolving features at 25–100 μm at production speed. The challenge with calendered surfaces is gloss. The same compaction that makes the electrode dense also makes it reflective, which creates noise that traditional rule-based vision struggles with.

AI vision is where the accuracy gap closes. Schoo et al. compared a 44-feature traditional ML classifier against ResNet-18 on the same 820-image dataset across eight defect classes. Feature-based ML hit 95.0% reclassification accuracy. ResNet-18 hit 96.3%. Choudhary et al.^[10] deployed YOLOv5 on a calendering machine running at 9 m/s with precision and recall that the authors describe as outperforming any human in the loop. Zhou et al.^[11] reported +2.4% precision, +2.3% recall, and +1.4% mAP over baseline YOLOv8 using a modified architecture with Ghost convolution and coordinate attention. The honest comparison: manual inspection sits at 70–80% accuracy with high inspector-to-inspector variance. Published AI systems on calendering data sit at 95–98% accuracy on their training distribution.

Where most manufacturers get this wrong

The defect detection conversation usually focuses on the camera. That's the wrong fixation. The real problem on a production line isn't pixel resolution. It's the lag between process drift and defect appearance.

What you actually see on the floor is this. The calender runs nominally for two shifts. Thickness drifts by 1.5 μm over 90 minutes. Nothing trips a spec limit. Particle fracture is happening, but it's invisible because nano-XCT isn't an inline tool. Coating gloss varies slightly, which the brightfield camera registers as noise rather than signal. Three days later, formation testing shows capacity fade on a batch of cells, and root-cause analysis gets stuck because the electrode is already in a can.

Published research evaluates models on curated datasets where defects are well-defined. The Schoo et al. 95–96% number is real, but it's measured on 820 hand-labeled examples across eight classes. Production data has long-tail defect types that don't appear in any training set. It also has surface gloss variation, lighting drift, web vibration, and operator-dependent thresholds that destroy the assumption of clean inputs.

The gap that matters is the lag between process parameter drift and downstream defect manifestation. Faraji Niri et al.^[14] used ML on pilot-line data to quantitatively link calendering gap, temperature, and porosity to impedance and capacity fade. Most production lines treat calendering as a thickness-control problem when it's actually a multi-variable porosity-and-microstructure problem.

What AI detection actually changes about this problem

AI vision changes three things specifically.

First, it closes the resolution gap on known defect classes. CNN backbones like ResNet-18 and ResNet-50, used by Schoo et al.^[9] and Badmos et al.^[12], reliably classify pinholes, agglomerates, mud cracks, scratches, and contamination at 95–98% accuracy on labeled data. YOLOv5, YOLOv7, and YOLOv8 variants handle real-time bounding-box detection at production line speeds. Mattern et al.^[13] showed transformer-based detectors outperforming CNNs on accuracy with an inference-speed penalty, which matters when choosing between edge deployment and server-side inference.

Second, it shifts inspection from sampling to continuous coverage. A human inspector covers maybe 5% of an electrode web. A line-scan AI vision system covers 100% at 60–100 m/min. That changes the statistics of what you catch.

What AI vision still can't do well without additional process context: detect subsurface particle fracture during calendering, predict defects from parameter drift before they appear visually, and generalize to defect types absent from training data. Novel defect modes, the kind that emerge during chemistry changes or new electrode formulations, are exactly what a pure vision system misses.

How Niobia AI's approach is different

We combine vision-based defect classification with process-parameter telemetry from the calender, so the model learns the relationship between roll gap drift, temperature variance, line speed, and the defect signature that appears 30–90 minutes later. The goal is to detect drift before it becomes scrap, not just count scrap after it happens.

For root cause analysis specifically, Niobia AI reduces the time from defect detection to confirmed root cause by 50×. A manual RCA on a calendering defect batch, pulling shift logs, correlating process data, reviewing inspection images, and tracing back to the upstream coater, typically takes 3–5 days. Our platform runs that analysis continuously in the background and surfaces a structured root cause report within minutes of a defect cluster appearing. The report includes defect images, correlated process parameters, probable root cause ranked by confidence, and recommended corrective actions.

When a defect type appears that doesn't match any existing category, Niobia AI automatically catalogues it as a new entry in your facility's defect library. That includes the visual signature, the process context at time of occurrence, and the cell-level outcome if formation data is linked. Over time, this builds a proprietary defect taxonomy specific to your chemistry, line configuration, and operating conditions, which is exactly what generic vision models can't give you.

Summary

Calendering defects in lithium-ion electrode manufacturing fall into six categories: edge cracking, delamination, pinholes, thickness non-uniformity, wrinkling, and active material fracture, with each driven by some combination of roll pressure (typically 60–200 MPa apparent for NMC), roll temperature (25–145 °C), line speed (60–100 m/min), and incoming coating uniformity. Manual inspection catches roughly 70–80% of defects. AI vision systems using ResNet, YOLO, and transformer architectures hit 95–98% accuracy on published benchmarks, and cover 100% of the web versus the 5% that human sampling achieves. The economic stake is clear: calendering is 5.19% of cell manufacturing cost, but a defect missed here propagates into formation and aging at 32.6% of cost, with documented scrap costs of roughly €30,000 per day per 1% scrap on a 40 GWh line. Niobia AI addresses this by combining inline defect detection with process-parameter correlation and automated RCA, cutting root cause investigation time by 50× and building a facility-specific defect catalogue that improves with every production run.

About the author

Dr. Gaurav Jha is the Founder of Niobia AI, which builds AI-powered defect detection and process intelligence platforms for battery gigafactories. His PhD focused on fast-charging niobium pentoxide (Nb₂O₅) based nanostructured anodes, with broader research spanning gas sensors, ion sensors, and energy storage materials. At Intel, he worked on wet etch defect reduction in 5nm and 7nm chip fabrication, developing a hands-on instinct for process root cause analysis at scale that translates directly to electrode manufacturing.

He returned to batteries to develop one of the first large-scale lithium-sulfur cathode coatings at Lyten, then moved to Sila Nanotechnology where he worked on silicon anode particles for high energy density and fast-charging applications across consumer electronics and automotive programs. Across these roles, Dr. Jha led manufacturing scaleup from lab to high-volume production, conducted industrial root cause investigations, commercialized key materials products, and developed new electrode chemistries from first principles. He founded Niobia AI to bring that depth of manufacturing and materials science experience into an AI platform built specifically for the production floor.

References

1. VDMA & PEM RWTH Aachen. (2021). Production Process of a Lithium-Ion Battery Cell. Industry Guide, VDMA Battery Production. https://vdma-industryguide.com
2. Meyer, C., Bockholt, H., Haselrieder, W., & Kwade, A. (2017). Characterization of the calendering process for compaction of electrodes for lithium-ion batteries. Journal of Materials Processing Technology, 249, 172–182. https://doi.org/10.1016/j.jmatprotec.2017.05.031
3. Wheatcroft, L., et al. (2024). In Situ Fracture Behavior of Single Crystal LiNi₀.₈Mn₀.₁Co₀.₁O₂ (NMC811). Batteries & Supercaps, 7(7), e202400077. https://doi.org/10.1002/batt.202400077
4. Günther, T., Schreiner, D., Metkar, A., Meyer, C., Kwade, A., & Reinhart, G. (2020). Classification of Calendering-Induced Electrode Defects and Their Influence on Subsequent Processes of Lithium-Ion Battery Production. Energy Technology, 8(2), 1900026. https://doi.org/10.1002/ente.201900026
5. Schreiner, D., et al. (2023). Simulation of the Calendering Process of NMC-622 Cathodes for Lithium-Ion Batteries. Energy Technology, 11(3), 2200442. https://doi.org/10.1002/ente.202200442
6. Ngandjong, A. C., Lombardo, T., Primo, E. N., & Franco, A. A. (2021). Investigating electrode calendering and its impact on electrochemical performance by means of a new discrete element method model. Journal of Power Sources, 485, 229320. https://doi.org/10.1016/j.jpowsour.2020.229320
7. Hidalgo, M. F. V., et al. (2023). Data of physical and electrochemical characteristics of calendered NMC622 electrodes and lithium-ion cells at pilot-plant battery manufacturing. Journal of Power Sources, 573, 233091. https://doi.org/10.1016/j.jpowsour.2023.233091
8. See, J. E. (2015). Visual Inspection Reliability for Precision Manufactured Parts. Human Factors, 57(8), 1427–1442. https://doi.org/10.1177/0018720815602553
9. Schoo, A., Moschner, B., Hülsmann, J., & Kwade, A. (2023). Coating Defects of Lithium-Ion Battery Electrodes and Their Inline Detection and Tracking. Batteries, 9(2), 111. https://doi.org/10.3390/batteries9020111
10. Choudhary, A., et al. (2022). Autonomous Visual Detection of Defects from Battery Electrode Manufacturing. Advanced Intelligent Systems, 5(2), 2200142. https://doi.org/10.1002/aisy.202200142
11. Zhou, Y., Yu, X., Wang, Z., & Hu, J. (2024). A YOLOv8-Based Approach for Real-Time Lithium-Ion Battery Electrode Defect Detection with High Accuracy. Electronics, 13(1), 173. https://doi.org/10.3390/electronics13010173
12. Badmos, O., Kopp, A., Bernthaler, T., & Schneider, G. (2020). Image-based defect detection in lithium-ion battery electrode using convolutional neural networks. Journal of Intelligent Manufacturing, 31(4), 885–897. https://doi.org/10.1007/s10845-019-01484-x
13. Mattern, M., et al. (2025). A comparison of transformer and CNN-based object detection models for surface defects on Li-Ion battery electrodes. Journal of Energy Storage, 105, 114378. https://doi.org/10.1016/j.est.2024.114378
14. Faraji Niri, M., et al. (2022). Quantifying Key Factors for Optimised Manufacturing of Li-Ion Battery Anode and Cathode via Machine Learning. Energy Technology, 10(2), 2200893. https://doi.org/10.1002/ente.202200893
15. Liu, Y., Zhang, R., Wang, J., & Wang, Y. (2021). Current and future lithium-ion battery manufacturing. iScience, 24(4), 102332. https://doi.org/10.1016/j.isci.2021.102332
16. Bockey, G., & Heimes, H. H. (2024). Mastering Ramp-up of Battery Production. Whitepaper, Fraunhofer Forschungsfertigung Batteriezelle (FFB) / PEM RWTH Aachen. https://www.ffb.fraunhofer.de