Cross-batch RCA
Pull SPC, materials and electrochemistry into one hypothesis.
Cross-batch RCA traces a defect across many batches instead of inside one: line up SPC, XRD, EIS, and yield data batch by batch, find which batches show the problem and which don't, and the correlation tells you what changed and when.
What it measures
When a defect appears in some batches and not others, the batch axis itself becomes the experiment. The method is to put every data source on that same axis and read across:
- Process data per batch: SPC charts of coating weight, thickness, line speed, oven temperatures. Which batches ran differently, and on which parameter?
- Materials data per batch: XRD on retained powder samples, particle size, surface analysis. Did the material itself change between the good batches and the bad ones?
- Electrochemical outcomes per batch: capacity, coulombic efficiency, EIS signatures from cells built out of each batch. How does the defect actually express itself, and in exactly which batches?
The output is a correlation with a mechanism: bad batches share a fingerprint that good batches lack, the fingerprint points at a change (a material lot, a process shift, a date), and the change explains the defect physically. Good batches matter as much as bad ones; they are the control group that turns a story into a test.
How to read the output
Build one table or one set of aligned charts: batches in order on the x-axis, every signal as a row. Then look for the step. A defect that starts at batch 41 and an XRD lattice shift that starts at batch 41 is a lead; add an EIS charge-transfer rise in cells from batches 41 onward and the correlation is strong. Respect the time lags: cells tested this week came from electrodes coated last week and powder received last month, so the batch axis has to carry genealogy, not calendar dates. And demand discrimination before acting: the candidate cause must be present in the bad batches and absent in the good ones. A change that appears in both explains neither.
A real use case
A capacity low-tail shows up in finished cells, intermittently, over several weeks. Inside any single batch the investigation stalls: the cells pass incoming checks and the line ran in spec. Across batches the picture sharpens. SPC on coating weight is flat through the whole window, which clears the coating room. XRD on retained cathode powder shows a small lattice-parameter shift that starts with one specific powder lot and persists for exactly two lots. EIS on the low-tail cells shows elevated charge-transfer resistance, and every low-tail cell traces back to electrodes made from those same two lots. Good batches before and after show none of it. The correlation has a mechanism (the supplier had changed precursor source, altering the surface), discriminates cleanly between good and bad batches, and survives the test: cells from the implicated lots on a second line show the same EIS signature. Three tools, one batch axis, one answer.
Common mistakes
- Investigating one bad batch in isolation. The information is in the contrast between good and bad batches; without the comparison there is nothing to correlate.
- Correlation without a mechanism. Batch data is full of coincidences; a correlation only becomes a cause when a physical pathway connects the change to the defect.
- Ignoring time lags and genealogy. Lining up this week's cell results with this week's coating data misses the real pairing by weeks.
- Only characterizing the failures. The good batches are the control group; skipping them is how teams convict an innocent variable.
- Letting each data silo conclude separately. Three teams each seeing "nothing significant" in their own tool can be one cause spread across all three.
Three lenses, one batch axis
Niobia runs each lens natively: SPC-style trend review on process data, XRD and particle-size analysis on retained material samples, and capacity, efficiency, and EIS analysis on the cells each batch produced. Inside an investigation it pulls those results onto one evidence matrix, batch by batch, with the hypothesis each signal supports or weakens, and its triage explicitly weighs recurrence so a signature that has appeared before is matched to its earlier investigation instead of starting from zero. The honest scope note: fully automated correlation mining across a plant's complete batch genealogy is partially supported today. The fusion happens inside guided investigations on the batches and data brought into them, and Niobia states that boundary rather than claiming a push-button genealogy miner.
Frequently asked
What data do I need before cross-batch analysis is possible?
Batch identity carried through the process (which powder lots went into which electrodes into which cells), at least one measurement per batch from each lens you want to use, and both good and bad batches in the dataset. Genealogy is the non-negotiable piece: without lineage, batch correlations are guesses.
How many batches does it take to see a pattern?
Enough to contain the transition and controls on both sides: several good batches, the first bad one, and what follows. A defect that starts cleanly at a batch boundary can be localized with a handful; an intermittent defect needs more, because the discriminating test is statistical rather than visual.
Why do single-batch investigations miss these causes?
Because inside one batch everything is consistent: the material, the settings, the date. The cause of a batch-to-batch difference is only visible in data that spans batches, which is also why it hides from teams whose tools each show one batch at a time.