Catching Scrap Early in Battery Gigafactories

A battery gigafactory is sold on a single number: nameplate capacity in gigawatt-hours per year. The number that decides whether the plant makes money is the one nobody puts on the press release — the share of cells that actually pass. In the first years of a new line, that gap is brutal. Fraunhofer's research institute for battery cell production puts scrap rates of 15 to 30 percent as common in the early life of a cell plant, and reports that even after five years reject rates often sit around ten percent (Fraunhofer FFB, 2024).

That is not a rounding error. It's the difference between a factory that returns capital and one that bleeds it. And the maddening part, for anyone who has stood on a coating line, is that the cell scrapped at the end of the process was usually spoiled near the beginning — in a coating lane, on a calender, in a slurry batch — hours and hundreds of metres of web before anyone measured it. Quality in cell making isn't an inspection booth at the gate. It's a data problem strung along a roll-to-roll line, and most plants are still trying to solve it after the fact.

What a point of scrap actually costs

Start with the money, because it reframes everything else. Fraunhofer's analysis of ramp-up economics is blunt about the arithmetic: each percentage point of reject rate costs roughly €30,000 per day and about €10 million per year, so a 30 percent reject rate at full capacity runs to around €900,000 per day (Fraunhofer FFB, 2024). Those numbers feel large until you see where cell cost comes from.

In a process-based cost model for GWh-scale production published in Communications Engineering, materials account for around 78 percent of total cell cost for an LFP chemistry and over 82 percent for NMC811 (Communications Engineering, 2024). The same model lands full cost in the range of roughly $101 to $109 per kWh for NMC811 and $105 to $113 per kWh for LFP at a 10 GWh plant in Germany. The headline there is not the absolute figure. It's the split: a cell is mostly purchased material with a thin skin of conversion cost on top.

So a scrapped cell is, overwhelmingly, thrown-away cathode and anode material that was already paid for. You don't recover most of it. You don't recover the energy spent drying the coating or the cleanroom hours that went into it. This is why scrap in cell making hurts more than scrap in a labour-heavy process: the value is locked into the material early, and every metre of bad electrode carries that value into the bin. Cutting reject rate isn't a quality nicety. On a materials-dominated bill, it's the cheapest cost lever on the floor.

The market backdrop sharpens it again. The IEA reported that NMC cells were less than 25 percent more expensive than their LFP equivalents in 2023, down from a 50 percent premium in 2021, and that producing a cell in the United States is nearly 20 percent more expensive than in China even before regional material differences (IEA, Global EV Outlook 2024). When margins between chemistries and regions are measured in low tens of percent, a yield gap of the same size is the whole game.

Why the defect stays hidden until it's expensive

If scrap is so costly, why does it persist at double digits years into a ramp? Because the way electrodes are made fights against catching problems early. Three mechanisms keep the defect hidden.

The process is continuous and fast. Electrodes are made roll-to-roll: slurry is coated onto metal foil, dried, calendered, and slit, with the web moving continuously. The US Department of Energy's assessment of roll-to-roll processing notes that quality control on moving webs is hard precisely because the measurement has to be made while the material is in motion, at production speed (US DOE, Quadrennial Technology Review, 2015). And a coating defect that starts in one lane keeps printing, metre after metre, until something downstream flags it, if anything does.

Most checks happen offline, after the fact. A recent review of process control for electrode manufacturing makes the gap explicit: the key electrode properties that govern cell performance — areal mass loading, coating thickness, and porosity — are still largely characterised offline, on samples, rather than measured and controlled in real time across the full web (arXiv preprint, 2025). Offline sampling tells you what a coupon looked like an hour ago. But it doesn't tell you that lane three has been drifting heavy for the last 400 metres.

The worst defects are latent. A coating deviation rarely announces itself. A mass-loading error or a sub-visible inclusion produces a cell that looks fine, assembles fine, and only reveals itself in formation, in capacity grading, or worse, in field aging. But by the time the electrical test at end of line fails the cell, the cause is buried hundreds of metres upstream and possibly days in the past. The signal that would have explained it, the gauge trace, the camera frame, the oven profile, usually sits in a different system from the test result, if it was kept at all.

And the clock makes it worse. Formation and aging hold cells for days to weeks before they grade, so a coating problem made on Monday may not show up in grading data until the following week. By then the line has produced kilometres more of the same electrode at the same setpoint, and all of that work-in-progress sits between the cause and the detection, suspect until proven otherwise. A feedback loop that ought to close in minutes instead closes in weeks. That lag is the real defect; the streak in the coating is just what triggers it.

The defect taxonomy itself is well understood. Inline studies of coated electrodes catalogue the recurring failure modes, and where each one comes from matters for catching it.

Machine-vision work on electrode coatings classifies essentially this set — scratches, bubbles, metal leakage, particles, and de-carbonised regions — from line-scan imagery (Coatings, MDPI, 2024). But metal inclusions are the one to lose sleep over: a stray conductive particle in the wrong layer is a latent internal-short hazard that no amount of downstream electrical screening reliably catches. That's a product-safety problem, and it lives or dies on detection at the coating step.

Put the three mechanisms together and you get the structural reason scrap stays high: the line generates the evidence of a defect in real time but doesn't act on it in real time. The PLC controlling the coater knows the pump speed. The thickness gauge knows the caliper. The camera sees the streak. The MES knows the lot. Formation knows the cell failed. None of them are talking to each other on a timescale that lets you stop making bad electrode.

Fix one: measure on the web, not at the gate

The first move is to push measurement upstream and inline, onto the moving web, so a deviation is caught in metres rather than rolls. Two technologies do the work.

Inline optical inspection is the mature piece. Line-scan cameras with controlled lighting scan the coated surface continuously and flag brightness anomalies — the streaks, pinholes, and inclusions in the table above — at web speed, with the defect's position recorded along the roll for later tracking (Batteries, MDPI, 2023). The classification layer on top has matured from hand-tuned thresholds to machine-vision models that separate defect types and cut false positives (Coatings, MDPI, 2024). But the point of the camera isn't a pretty defect map. It's that a metal inclusion gets caught at the coater, before it's wound into a roll and sent to assembly.

Inline metrology is the less-mature piece, and the bigger prize. Coating thickness and areal mass loading are the properties that drive capacity and consistency, and they're exactly the ones still measured offline in most plants. Closing that gap, measuring mass loading, thickness, and porosity across the web and feeding it back to the coater, is the explicit opportunity flagged in the current process-control literature (arXiv preprint, 2025). The measurements have to be traceable to be trusted; gauges drift, and a quality decision is only as good as the calibration behind it, which is where NIST-traceable measurement standards earn their keep (NIST).

None of this is exotic hardware. Cameras, beta or x-ray basis-weight gauges, and laser caliper sensors are off-the-shelf, and most plants already own a fair share of them. The hard part is almost never the sensor. It's getting the sensor's output off the line, time-aligned with everything else the line is producing, and in front of something that can act on it before the next roll is wound. A gauge that logs to its own proprietary file, read once a shift, may as well not be there.

Fix two: close the loop between the line and the model

An inline gauge that streams to a chart on a panel nobody watches has changed nothing. The defect stays hidden because the data stays trapped — in the PLC, in the camera's own software, in a historian that no model ever reads. The fix is to treat the line as one instrumented system and connect the layers that currently sit in silos.

The reference for those layers is ISA-95: sensors and PLCs at the bottom, the manufacturing execution system at Level 3 coordinating orders and lots, the enterprise above. Quality lives in the seam between them. A coating-weight excursion is a Level 1/2 signal; the decision to quarantine a roll is a Level 3 action; the cost of the scrap shows up at the enterprise. If those layers only exchange data in nightly batches, the loop is too slow to stop bad product.

Two things have to be built across that seam.

Edge telemetry with the decision close to the line. Statistical process control on coating weight, on caliper, on defect counts per metre should run at the edge, in seconds, where it can raise an alarm or hold a roll — not after a round trip to the cloud. The cloud is for the model and the history; the edge is for the reflex. That layer — edge telemetry that turns sensor streams into decisions on the line — keeps the heavier analytics and model training off the critical path.

Cell genealogy keyed to position. Every cell needs a traceable record back to the electrode it was made from, and every metre of electrode needs to carry its inline data — coating weight, defect map, oven profile — indexed by position on the web. When a cell fails formation or grades low, you want to walk the genealogy straight back to the coating lane and the timestamp that produced it. Without that linkage, end-of-line failures are just a tax; with it, they're a training signal. This is the practical content of a quality management system under ISO 9001:2015 — not the binder of procedures, but the actual traceability and corrective-action loop the standard asks for.

Once genealogy exists, the model becomes possible. Correlate end-of-line and formation results against the upstream inline signals, and the latent defects start to surface their causes: this band of coating-weight drift predicts that capacity scatter; this oven excursion precedes those low graders. That correlation is the whole reason to instrument the line. And a model trained on data that never leaves the PLC is a model that never gets trained.

A sequence that pays back

You don't boil the ocean. The order of work matters more than the size of the budget, because each step has to earn the next. A pragmatic sequence:

1. Instrument the coater and calender first. Inline basis-weight gauge, caliper, and line-scan camera on the coating step — that's where most of the recoverable scrap is born and where material cost is committed.
2. Put everything on one clock. Time-align camera frames, gauge traces, and PLC tags to a common timebase and tie them to web position. Unsynchronised data can't be correlated later, and correlation is the point.
3. Run SPC at the edge. Detect drift in seconds and act — hold the roll, flag the lane — before bad electrode keeps printing. Don't wait for a nightly report.
4. Build genealogy before you build models. Link every cell to its electrode and its inline data. The data infrastructure has to exist before any learning model has something honest to learn from.
5. Close the analytics loop, then the control loop. First correlate failures to upstream signals to find root causes. Only then, and cautiously, feed measurements back to setpoints — automated control of a coater is a step you take once the measurement is proven and traceable.

The payback follows the cost structure. Because materials are roughly four-fifths of cell cost (Communications Engineering, 2024), every point of scrap you prevent at the coater is mostly recovered material, and the recovery compounds across a 10 or 20 GWh line. But against Fraunhofer's €10 million per percentage point per year, the instrumentation to catch a metal inclusion or a coating-weight drift before it becomes a scrapped cell pays for itself in a timeframe measured in weeks, not capital-plan years.

For teams standing up or stabilising a line, the deployment pattern is the same one that works across processing plants: instrument the step that commits the most value, get the data off the line and onto a common clock, and close the loop from measurement to decision. That's the substance of industrial AI deployment that survives contact with a real plant floor, as opposed to a dashboard that looks good in a demo.

The short version

Gigafactory scrap stays high not because the defects are mysterious but because the line records them in one place and decides in another, too late to matter. The defects are catalogued. The cameras and gauges exist. The cost case is overwhelming: on a bill that's four-fifths material, yield is the cheapest lever there is. What's usually missing is the plumbing: inline measurement on the moving web, time-aligned data off the line at the edge, genealogy that links a failed cell back to the metre of electrode that doomed it, and a model fed by all of it. Build that, and the defect that used to surface at the gate surfaces at the coater instead, where you can still do something about it.