Why do defects occur in battery production? Explainable AI supports process optimization.

Artificial intelligence is increasingly reshaping how we work, live, and think. But it’s often unclear just how AI arrives at its results. Why does ChatGPT suggest using five eggs for this cake when every other recipe for it says to use four? Why does AI hallucinate events that never happened? What logic does AI follow to make decisions? The good news is that explainable AI is on the rise, at least when it comes to process optimization for production chains. In the future, this will include battery production. This is in part thanks to the efforts of Thomas Stocker, group manager for production management at the Development Center X-ray Technology. He uses this technology to find defects before they can cause any damage. The aim is to achieve a consistent level of battery quality while also reducing waste and improving safety. In this interview, you will discover how explainable AI works, how it can be used, and what process optimization has in store for us in the future.

Mr. Stocker, explainable AI is a relevant factor in numerous fields. In our current series on batteries, we’re focusing mostly on sustainable electromobility. What role does explainable AI play here?

Two of the key considerations when producing batteries for electric vehicles are waste reduction and quality assurance. Unfortunately, the scrap rate in battery production is still rather high. For certain production cycles, it can be as much as 20 percent – in other words, one in five batteries must be discarded. In addition to cutting down on waste, significantly lowering this scrap rate would lead to reductions in the amount of raw materials and electricity required for production. That’s a more efficient approach than having to start recycling once production is underway. In addition, explainable AI can help counteract the shortage of skilled workers by performing many tasks autonomously. Sometimes, there’s only one expert for a given system, and they end up working at a different location every day. As our customers tell us, this leads to more complex workflows as well as delays. But in the end, AI systems notice anomalies and correlations that a human might well miss.

A technology that offers so many benefits must be complex. How does explainable AI work?

I like to illustrate that with the example of neural networks, which we use a lot in our work. We train these networks by feeding them mountains of data. But once training is complete, what actually happens within the network and why it reaches a certain decision is something for explainable AI to figure out.

Let’s say we’ve trained the network using pictures of cats and dogs. Once trained, it can tell the two animals apart. But precisely how the network decides if an image is a cat or a dog is usually unclear.

Explainable AI tries to work backwards to understand the network’s decision-making process. The difference is that the AI’s decision logic is transparent. What areas of the image tipped the balance? How were these areas processed?

For a project involving foundries, for example, we trained a neural network to identify defect sizes in cast products. In addition to spotting the defects themselves, we also wanted to discover how they happened in the first place. We checked to see whether there were process parameters in the production of cast parts that could affect defect size.

The network was able to find a correlation between specific production parameters and a certain defect size. In other words, explainable AI was able to help identify which of these process parameters were responsible. As a result, we could tell the customer which process parameters to adjust to prevent certain defects.

Customers often fear that implementing this kind of technology will involve a protracted and problematic transition. Is this the case with explainable AI for process optimization?

The first condition is having integrated an imaging method, such as computed tomography or X-ray technology, into the production process. The second is the ability to trace parts – in other words, to know which machine made the part that is to be examined. If these conditions have been met, we can integrate explainable AI quickly and easily into the existing production process.

When it comes to the type of AI, there are two main options. One is to train the AI in advance; once that’s complete, the system stops learning. This type delivers consistent results and is easier to monitor. The other option is to allow the system to continue learning, which means it will benefit from new data generated during production and keep on improving. But this also means it has to be monitored more closely. Both approaches are generally suitable for quality assurance in battery production. One advantage of this technology is that it doesn’t require all that much hardware. Balanced against the savings from waste reduction, the costs are manageable.

What does the future have in store for this technology?

Going forward, we want to redesign our AI to make it easy for anyone to use. Our long-term vision is for the AI to autonomously regulate itself and eliminate defects automatically. We’ve already achieved an extremely wide range of applications. No matter whether our technology is used in battery production, the manufacture of cast parts, or in other production processes, it finds the cause of defects. In this way, it helps create products that are better and more sustainable.

Mr. Stocker, thank you for talking to us today.

Article by Lucas Westermann, editor, Fraunhofer IIS Magazine