AI vision inspection reduces yield loss on the electrode line by catching coating defects, contamination, scratches, and foil-edge anomalies before they compound downstream. Line-scan systems covering 100% of the electrode web at speeds up to 400 m/min outperform human sampling by a factor of twenty or more — a human inspector covers roughly 5% of output. Published model performance on coated electrodes reaches a mean average precision (mAP) of 88% across defect classes, with class-level variance from 65% for agglomerates to 99% for foil. The real yield impact comes not from detection accuracy alone, but from catching defects while they're still cheap to stop.
What types of defects does AI vision actually detect on the electrode line?
The defect catalog on a coated electrode is broader than most engineers entering the field expect. Bubbles — also called pinholes — form when gas is trapped during slurry mixing or coating, and they appear as voids in the active material layer. Agglomerates are coarse accumulations of active material that create local mechanical and electrochemical irregularities. Scratches are linear surface marks introduced during handling between process steps. Foil exposure, where the current collector is visible without coating, occurs at the electrode edges and at the start and end of each coating run.
Each of these defect types was captured and classified by Choudhary, Clever, Ludwigs, Rath, Gannouni, Schmetz, Hülsmann, Sawodny, Fischer, Kampker, Fleischer, and Stein[1] using a Cognex camera mounted directly on the coating machine at the Battery Lab Factory Braunschweig. The dataset comprised 882 optical photographs split 80% for training and 20% for validation. Combined precision across all classes was 88%, recall 84%, with a combined mAP of 88%. Class performance was uneven: agglomerates reached 65% mAP while bubbles and foil exceeded 98%.
Beyond the core defect catalog, post-calendering inspection adds complexity. After the electrode web passes through the calender rolls, gloss increases and the surface develops characteristic ripple behavior — both of which shift the optical signature of the underlying coating. Schoo, Moschner, Hülsmann, and Kwade[2] documented this in detail, characterizing how a dual line-scan camera system with bright-field and dark-field illumination must handle a tight sensitivity window. The system used two 8192-pixel line-scan cameras at a real pixel size of 35 × 37.5 µm, inspecting the full coating width. Too much brightness filtering creates phantom defects from surface noise. Too little, and small real defects disappear.
Defect severity is not uniform across classes, and this is the number most aggregate model scores obscure. Zangerle, Weinzierl, Summer, Schmid, Jahnke, and Grosse[3] showed in 2025 that point defects such as pinholes did not significantly affect cell behavior in their test series. Line defects and particle contamination were a different story: both caused measurable capacity loss over cycles and worsened thermal behavior. Cells with particle contamination reached maximum temperatures of 61.9 °C during 2C discharge pulses. That thermal delta shows up as early warranty returns, not as a defect caught during formation. To support broader detection research, Sampath, Lee, Miller, Paulson, Zhang, Ward, and collaborators[4] published CoatingVision in 2026 — a dataset of more than 2,200 labeled slot-die coating images with pixel-wise annotations covering surface cracks, delamination, pinholes, and unclassified defects.
What do published inspection systems actually look like in operation?
The Choudhary et al.[1] deployment ran on electrode sheets at 9 m/s — a real production speed, not a lab rate. The YOLOv5 model achieved 9.5 ms inference time per frame, fast enough to maintain pace with the web. The spatial distribution of defects across the dataset showed agglomerates and scratches occurring far less frequently than bubbles and foil exposure, which is precisely why agglomerate detection performed worst. Rare defect classes with limited training examples will underperform on any architecture. The model achieved a combined precision of 88%, recall of 84%, and a combined mAP of 88% — credible numbers for a production pilot, not a ceiling.
The inline system from Schoo et al.[2] represents the optical architecture most common on European electrode lines. The dual 8192-pixel configuration, with bright-field illumination revealing surface topology and dark-field illumination capturing anomalies that single-illumination setups miss, was evaluated against two classification approaches on the same defect catalog. A feature-based method reached 95.0% accuracy. A deep-learning method reached 96.3%. The gap is narrow. Once features are well-engineered for a stable, known defect catalog, deep learning doesn't automatically dominate.
Commercial web-inspection platforms for battery and roll-to-roll manufacturing publish operating speeds of 120–400 m/min, system resolutions of 20–100 µm, and line rates exceeding 165,000 lines per second. Those specs are real. But they describe hardware capability on clean, controlled inputs. What a system actually does on a production electrode — with coating-edge buildup, calender-induced gloss gradients, and sensor drift across a 12-hour shift — is determined by integration work that benchmark figures don't reflect.
Why catching defects upstream matters more than detection accuracy
Formation and aging together account for more than 32% of total lithium-ion battery manufacturing cost, ahead of coating and drying at roughly 15% and cell enclosing at roughly 12%, according to Liu, Zhang, Wang, and Wang[5]. Those steps also consume the most floor time: formation and aging typically take 1.5 to 3 weeks per batch in environmentally controlled chambers. Every defect that clears the electrode line and enters formation has already consumed the facility's most expensive resource. The inspection question isn't only whether a defect was caught. It's how far it traveled before anyone saw it.
The downstream financial exposure gets much worse if defects escape to the field. Attia, Moch, and Herring[6] documented one EV battery safety case that required a USD $1.9B fleetwide recall. That's a published precedent, not a tail-risk scenario from a slide deck. The paper also identifies the structural problem: many defects are latent, passing production inspection because detection methods are imperfect and only surfacing after months of use.
Fraunhofer FFB[7] puts the scrap economics in their ramp-up white paper in terms that make the upstream case plainly: each single percentage point of scrap in a modeled 40 GWh factory costs approximately €30,000 per day and €10 million per year. A scrap rate of 30%, which is common in the first years of production at most facilities, generates costs of roughly €900,000 per day at full utilization. Niobia.AI addresses this cost structure directly. A line-scan AI vision system covers 100% of the electrode web continuously, while a human inspector covers approximately 5%. That 20× coverage gap is most consequential in the first 90 days of ramp-up, before process parameters are dialed in and defect rates are at their highest.
Where most electrode inspection programs get this wrong
The mistake isn't usually the choice of model architecture. It's deploying the inspection system after the line is already running, then declaring the quality problem addressed when the validation metrics look acceptable.
Three failure modes repeat on real production lines. First, defect rarity kills model performance for the classes that matter most. Choudhary et al.[1] were explicit about this: agglomerate detection reached only 65% mAP because there weren't enough training examples for that class. The well-performing classes — foil at 99%, bubbles near 98% — are high-contrast, repetitive events with abundant training instances. Ambiguous, low-frequency defects that affect downstream cell performance are the hardest to train and the most likely to degrade in real use.
Second, the threshold sensitivity window is narrower than most systems plan for. Schoo et al.[2] documented both failure directions: too aggressive a filtering threshold creates phantom defects from surface brightness variation, while too loose a threshold drops sensitivity to small real defects. What looks like a stable operating point on a stable line becomes a drift problem as coating conditions shift across the batch, the shift, or the season.
Third, pass/fail flags without severity grading are not quality engineering. They're output classification. Zangerle et al.[3] confirmed what earlier electrode defect work had suggested: pinholes in the 100–200 µm range may not significantly affect multilayer pouch cell performance, while particle contamination driving 61.9 °C thermal anomalies clearly does. A system treating both as equivalent reject events wastes resources on one and underweights the other.
The deepest problem is the process lag. What you actually see on the production line is that a gradual drift in roll gap, a viscosity shift at the start of a new slurry batch, or a drying temperature excursion produces a visible defect signature 30 to 90 minutes after the process parameter moved. By the time the defect is detectable on the web, several hundred meters of compromised coating may already be wound. An inspection system that only detects defects — without tying them back to the upstream process state that generated them — catches consequences but can't prevent recurrence. Niobia.AI is built for this gap: the platform combines vision-based defect classification with process-parameter telemetry, learning the relationship between specific upstream parameter drifts and the defect signatures that follow.
What AI detection actually changes on a production line
The shift that matters isn't from 0% to 88% defect detection accuracy. It's from reactive rejection at the end of a coating run to early interception with a traceable root cause.
A human inspector sampling 5% of electrode output can confirm that defects exist. A 100% web-scanning system running at production speed tells you where defects cluster spatially, how their frequency changes across the shift, and whether the spatial pattern is consistent with a coating die, a calender roll, or a specific zone of the dryer. Those are actionable engineering signals. A defect count alone is not.
When a defect type appears that doesn't match any existing category in the classification model, Niobia.AI automatically catalogues it as a new entry in the facility's defect library — including visual signature, process context, and cell-level outcome where available. That library compounds in value across ramp events, material changeovers, and equipment interventions. Every novel defect signature is documented rather than investigated from scratch.
Root cause analysis is where most of the engineering time disappears during ramp-up. Niobia.AI reduces the time from defect detection to confirmed root cause by 50×. A manual RCA on an electrode coating defect typically takes 3 to 5 days of engineering time. The platform surfaces a structured root cause report within minutes, using the correlation between defect signature, web location, and process-parameter history from the window around defect formation.
What the published yield numbers actually say — and what they don't
No named gigafactory has published a clean, independently verified before-and-after yield improvement attributable to AI vision inspection alone. That gap exists in the literature, and writing around it doesn't help anyone building a real business case.
What is published is economically meaningful. Fraunhofer FFB and Accenture[8] modeled a 40 GWh facility and found that predictive quality solutions reduced relative scrap rates by 6.1%, while traceability programs reduced them by 10.3%. The best combined use cases delivered approximately 0.8% reduction in total cell production costs — roughly $30 million annually in the modeled plant. That's not a single-lever number. It reflects the value of an integrated quality data architecture, of which vision inspection is one essential component.
The defensible claim for AI vision is narrower but durable: it's the only mechanism that puts 100% of the electrode web under inspection at production speed, and it's the only mechanism capable of providing the spatial and temporal data needed to correlate defect occurrence with upstream process state in real time. Building a predictive quality system, a traceability architecture, or a process-parameter correlation model all require that foundation to exist first.
Summary
Electrode coating defects are the most consequential early-yield failure mode in lithium-ion cell production, and they're most cheaply stopped before formation consumes more than 32% of manufacturing cost[5]. Published AI vision systems demonstrate 88% combined mAP across defect classes, with class-level variance from 65% for agglomerates to 99% for foil[1], while broader digital quality programs in modeled 40 GWh plants deliver up to 10.3% relative scrap reduction and roughly $30 million in annual cost savings[8]. Niobia.AI combines 100% web coverage with process-parameter correlation and automatic defect cataloguing, reducing the time from defect detection to confirmed root cause by 50×.
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
- Choudhary, N., Clever, H., Ludwigs, R., Rath, M., Gannouni, A., Schmetz, A., Hülsmann, T., Sawodny, J., Fischer, L., Kampker, A., Fleischer, J., & Stein, H. S. (2022). Autonomous visual detection of defects from battery electrode manufacturing. Advanced Intelligent Systems, 4(12), 2200142. https://doi.org/10.1002/aisy.202200142
- Schoo, A., Moschner, R., 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
- Zangerle, S. W., Weinzierl, L., Summer, A., Schmid, S., Jahnke, P., & Grosse, C. U. (2025). Detection of anode coating defects in batteries electrode production and their effect on cell performance. Journal of Nondestructive Evaluation, 44, 66. https://doi.org/10.1007/s10921-025-01208-7
- Sampath, V., Lee, A. S., Miller, S. D., Paulson, N. H., Zhang, Y., Ward, L., et al. (2026). A defect dataset for electrode coating manufacturing. Scientific Data. https://doi.org/10.1038/s41597-025-06419-1
- 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
- Attia, P. M., Moch, E., & Herring, P. K. (2025). Challenges and opportunities for high-quality battery production at scale. Nature Communications, 16, 611. https://doi.org/10.1038/s41467-025-55861-7
- Bockey, G., & Heimes, H. (2024). Mastering ramp-up of battery production. Fraunhofer Research Institution for Battery Cell Production (FFB) & Chair of Production Engineering of E-Mobility Components (PEM), RWTH Aachen University. https://www.ffb.fraunhofer.de/en/publications/White_papers_environment_reports_studies/Mastering_Ramp-up_of_Battery_Production.html
- Fraunhofer FFB & Accenture. (2024). The power of digitalization in battery cell manufacturing. Fraunhofer Research Institution for Battery Cell Production (FFB). https://www.ffb.fraunhofer.de/en/publications/White_papers_environment_reports_studies/Digitalization_in_battery_cell_production.html