How BMW and the University of Zagreb are increasing cell production knowledge
BMW's Christian Siedelhofer explains how artificial intelligence, developed with the University of Zagreb, is cutting setup times, detecting defects invisible to the human eye, and building a validation framework for battery cells
BMW does not build battery cells in series production. It never has, and – on current plans – it does not intend to start. Yet the company has spent years investing money, capacity and engineering talent into developing its own cells, running its own pilot production line, and now applying artificial intelligence across nearly every stage of the process. There is a strong case being made for this approach.
According to Christian Siedelhofer, Group Lead Technology Development, Battery Cells at BMW, the rationale comes down to control over knowledge rather than control over output. BMW's approach rests on three internal competence centres: The Battery Cell Competence Center (BCCC) in Munich, which handles research and development from material characterisation through to future cell concepts; the Cell Manufacturing Competence Center (CMCC) in Parsdorf, which industrialises that R&D at pilot scale; and a Cell Recycling Competence Center (CRCC), operational for a little over six months, which brings direct recycling from theory into practice.
If you change component A to B, then change the pump speed RPM or another, more deeply hidden machine parameter...It's a very complex environment and the AI helps in these specific steps to find new starting points which are closer to the optimum of our existing process parameters
Building and operating this infrastructure – without ever scaling up to a gigafactory – Has provided BMW with a wealth of knowledge and leverage it would not otherwise have. It allows the company to set specifications for external cell suppliers without dictating exactly how those specifications should be met. BMW defines a specification sheet, and suppliers must hit certain performance results, but the route they take to get there is their own. That structure, in turn, creates a feedback loop – BMW's expertise can, in turn, support suppliers in troubleshooting their own production issues, and BMW gains insight it could not generate without running cells of its own.
Cutting setup times with predictive models
The headline figure attached to BMW's AI collaboration with the University of Zagreb is a 50% reduction in material and time during individual process steps. Siedelhofer is careful to frame what that number actually represents. "The number is an example of what is possible in optimising a single process step out of many," he explains, and the results are currently being transferred from BMW's prototype build facility onto the pilot line — it has not yet been rolled out across the board.
Rather than a single AI system governing the entire battery cell production process, BMW is deploying multiple, discrete prediction models, each targeted at specific steps and trained on process and product parameters pulled directly from live operations.
The clearest example is the very first stage of cell production; mixing powders and liquids into a slurry. Because BMW frequently changes materials, solvents and active materials at its prototype build stage, the process parameters must be reset each time. "In the past, this was done by experience of operators to set the right process parameters," Siedelhofer says. Historically, if BMW had run an identical combination before, replicating past settings was straightforward. But new combinations are the norm, not the exception, and that is precisely where the AI models come in.
"If you change component A to B, then change the pump speed RPM or another, more deeply hidden machine parameter," he explains. "It's a very complex environment and the AI helps in these specific steps to find new starting points which are closer to the optimum of our existing process parameters" — covering variables such as mixing speed and temperature. Crucially, the AI doesn't eliminate physical testing; it shortens the distance between a cold start and an optimised process window, cutting the setup time needed whenever materials change.
That sensitivity to small variation doesn't disappear once a product reaches stable series production, either. Even when a supplier commits to one product long-term, "stable is not an absolute value," Siedelhofer notes. Materials carry tolerances — humidity levels included — and exterior conditions remain uncontrollable. "This is why we really set up this live data support to our operators, so they can use the data they're getting from the machine, from process, and from product to optimise and to keep the process in a stable process window."
Building a cell validation framework from scratch
BMW's decades of experience validating vehicles doesn't translate automatically to battery cells – a technology the company has only been working with intensively for a fraction of that time. Siedelhofer confirms that the cell validation framework is effectively being built from the ground up.
That gap is part of the justification for BMW's continued investment in cell development despite not manufacturing cells in series. Understanding the technology deeply enough to build a validation framework – and to specify supplier requirements with precision — requires the kind of data most vehicle manufacturers simply haven't needed to collect before. "It is important to get as much data out of your production and product as possible," he says, adding that BMW built this thinking into its prototype and pilot line from the outset, "so that we can make decisions that are data driven."
That matters more, he argues, is because battery cells aren't a technology BMW has worked with for a century, unlike combustion engines and vehicle assembly. "We have to catch up, and I think we are really on a good path because we are very data focused." But gathering that data is genuinely difficult. Cell production moves from powders and liquids, through a roll-to-roll electrode process, before converging into a one-piece flow only at the very end. Aligning data across that entire chain — tracing which batch of active material powder ended up in a specific finished cell — is significantly more complex than tracking a process with a single continuous flow from start to finish.
Extending AI beyond the cell
BMW's use of AI in battery development isn't confined to individual cells. It extends into high-voltage battery assembly, where roughly 960 individual cells are combined into a single pack. BMW is establishing five new plants worldwide, one in each major region, and – according to feedback from partners relayed during the conversation – the company collects more production data than any other manufacturer in this space, using AI agents and assistants to refine processes across the board.
Detecting the "worst enemy" of a battery cell
Quality inspection is where BMW's AI ambitions meet their hardest technical test. Towards the end of the production process, BMW uses computer tomography (CT) to assess finished cells – a well-established imaging technology in general terms, but one Siedelhofer describes as genuinely progressive when applied to battery cell production at this level of detail. The target is what he calls "the worst enemy of a battery cell" – a particle inside it, small and difficult to detect.
The challenge is that this isn't a simple pass/fail judgement. Siedelhofer draws a contrast with older manufacturing checks – inspecting whether a cable is properly clicked into place on a combustion engine assembly line, for instance, where a small window simply shows white or black, correct or incorrect. Identifying a particle inside a battery cell is far less binary. CT scans also introduce their own visual artefacts from the imaging process itself, which the AI must learn to distinguish from genuine defects.
Training a model to make that distinction reliably would, in an ideal world, require examples from millions of produced cells containing millions of real defects. BMW doesn't have anywhere near that volume at pilot scale, and buying in external data isn't a viable shortcut – different cell chemistries and designs would simply teach the AI to detect the wrong thing. Instead, BMW deliberately creates validation samples containing controlled defects for AI training purposes.
In some cases, this means physically implanting particles into cells before scanning them; in others, it means generating defects virtually within CT images. "You need one AI to virtually implement defects into your CT images, to then train the other AI to detect those defects," Siedelhofer explains.
"The threshold for replacing an established quality gate is intentionally high. Any new technology must outperform the state-of-the-art in at least one aspect while matching it in all others
Replacing – eventually – the voltage-hold test
The industry's current benchmark for a final "OK" verdict on a battery cell is the voltage hold method; measuring a cell's voltage, holding it in quarantine for a set period, measuring again, and checking whether it falls within the correct process window. BMW's AI-supported quality processes are being measured directly against that standard, and the bar is deliberately conservative.
"The threshold for replacing an established quality gate is intentionally high. Any new technology must outperform the state-of-the-art in at least one aspect while matching it in all others,” Siedelhofer says of any new method intended to succeed an established one.
So far, BMW's AI-based approach identifies every error the existing voltage-hold method finds – the minimum threshold for consideration. That doesn't mean an imminent changeover, however. Siedelhofer is candid that a wider dataset is still required before BMW would trust the new method to run alone, even at series-production scale. The likely path, in his view, is not a hard switch but a period of running in parallel; both methods operating side by side until production reaches a genuinely stable phase, at which point BMW could transition over.
There's a tangible payoff even before full replacement happens. Conventional quality gates, Siedelhofer notes, often carry a high false "not OK" rate because they judge a cell on a single measured value rather than the fuller picture of how it was actually produced. By incorporating more process history into its assessment, BMW's AI-supported quality gates can reduce that false rejection rate – meaning less unnecessary scrap.
Working at pilot rather than gigafactory scale also gives BMW options that large-volume cell manufacturers don't prioritise. Because BMW isn't producing at true mass-market volumes, the company has the bandwidth to think carefully about what happens to any scrap it does generate – one of the drivers behind establishing its direct recycling competence centre in the first place.
Scaling: The "plumbing" problem, not the models
Asked where the biggest technical obstacle lies in scaling these AI models from prototype into series production, Siedelhofer is unambiguous; it isn't the models themselves. "The challenge in scaling is before you do analytics and predictions – the plumbing of the data," he says. Connecting robots and equipment to a usable data lake is, in his view, the real prerequisite for any analytics model to function – and that infrastructure work looks different at every scale and in every application.
That principle extends across technologies as well as production volumes. AI models BMW has developed for improving laser weld quality in battery cells, for example, can be adapted for use in high-voltage battery production, since welding also occurs there. The underlying model doesn't need reinventing – only the data feeding it needs to be reconnected, or "plumbed," differently.
The Parsdorf facility is designed with that scalability in mind. Rather than running cheaper, non-representative equipment, BMW has installed genuine, series-like machinery at every process step – for instance, the winder used at Parsdorf is the same type used in true series production, just deployed in smaller numbers. That design choice gives BMW clear visibility into how each individual process step would behave at full scale, even though total cell output at Parsdorf remains far below gigafactory volumes – volumes, Siedelhofer stresses, are the only way battery cell production becomes genuinely economic.
Towards a manufacturing copilot
The University of Zagreb collaboration was, in Siedelhofer 's view, never intended to stay confined to battery cells. From the outset, the ambition was for the AI tools developed through the project to extend well beyond battery production and into a broader concept — "some call it really a manufacturing copilot," he says, drawing a comparison to familiar office productivity copilots, but built for a production line instead.
Battery cell manufacturing remains, in his words, "special and a challenge" in its own right. But the AI capability built to solve its specific problems is explicitly designed to be portable to other production lines across BMW – one reason, he notes, that his team stays closely connected to colleagues working on high-voltage batteries and vehicle assembly more broadly. For BMW, the payoff of investing in cells it will never build at scale may ultimately extend well beyond the battery itself.
