AI-Driven Battery Life Forecasting Enhances New Design Validation

The AI tool built at the University of Michigan draws on tests of earlier batteries to predict the performance of new designs. (Image: Xin Zou, University of Wisconsin-Madison)

Tech Briefs: What motivated you to start this project?

Ziyou Song: I have been working in both academia and industry, so I know there is a pain point regarding battery testing and validation. When we have a new battery design, even though it’s not that different from the previous one, we always do all the lab testing under similar conditions. That will cost at least six months and consume a lot of energy. So, if we have a new design that is not too different, we were wondering if we could use those previous datasets. Could we leverage the knowledge and data patterns from the previous design to predict the lifetime of our new design? That would save a lot of time and energy.

We are looking at this problem from a perspective of AI for science, which is an emerging field — scientific prediction is one of the most important problems. We are specifically focusing on the cost effectiveness of the machine learning method. How can we use the least possible data to achieve better or similar prediction performance.

Tech Briefs: How do you deal with the fact that the battery design is not the same as the previous one?

Song: For this question, we have to delve into our specific framework. We leveraged a concept from a study several decades ago to mimic how human beings perceive the world and gain knowledge. For example, we learn many things roughly by our observations, accurately by taking classes, or by asking others. Our framework leverages a teaching approach called discovery learning. When we have a new battery design, we want to map it to existing battery cells to see how we can use their datasets for our new design, so we can make a lifetime prediction without prolonged testing.

We test the new design to observe its initial parameters and how they evolve during the first several dozens of cycles. Then we try to map this pattern to previous batteries because we have the complete datasets for them. If we see similar patterns, we know our new design is similar to the earlier one. We can then leverage that knowledge, the data patterns, to predict the lifetime of our new battery.

Tech Briefs: You say that you use significant energy when you test the batteries. Is that because you test batteries under load?

Song: Yes, exactly. When we test our battery cells, specifically for the lab cycle testing, we have to charge and discharge many, many times, for example, 1,000 cycles, 2,000 cycles, which will consume a lot of energy. So, if you can do only 50 cycles, that can save a lot.

We not only validate over fewer cycles but also validate fewer cells. For example, previously, because you wanted to cover different operating conditions, you would have to use maybe 100 cells, 100 channels to test under different conditions. But now, we only need to cover some of them — for example, 10 cells, 10 operating conditions, 50 cycles. So, we save energy both from shorter cycling times and also by testing fewer battery cells.

Tech Briefs: What are some of the conditions you test for?

Song: Temperature is a very important metric. Another is how we use our batteries. For example, if you have a cell phone and I have a cell phone, we are probably using them in very different ways. I want to fully charge my phone all the time, so my cell phone battery degrades very quickly. But if you just use your cell phone in the middle of the state of charge (SOC) range, it will probably last longer. Since different people use a battery differently, we want to cover as many conditions as possible.

Tech Briefs: I read that there are three components of your system: learner, interpreter, and oracle. Would you describe those please?

Song: Those are the three components of our discovery learning. The first is the learner — the one that asks the question, that decides which cell designs and cell prototypes to test. The learner will pass the command to the interpreter, which will leverage the early cycling test data to build a feature space that consists of physical parameters, i.e., features, which can then be interpreted by a battery expert. So, we're not just using some meaningless statistical features; we're using features that can be interpreted by a battery engineer or scientist. Then the interpreter will provide these features to the oracle. The oracle will do so-called zero-shot learning, which just means it predicts the lifetime of new battery designs without additional experiments. The oracle will then pass its prediction results back to the learner, which will trust these results and learn from them. After a number of runs, the learner will be equipped to predict new battery designs by itself. And finally, the learner and the oracle will together make predictions for all the battery samples.

Tech Briefs: Professor Song, you said that you pick batteries to learn from that had similar patterns to what you're doing now. How do you choose which batteries to use as your models?

Song: It's primarily from the interpreter. For example, when we have the evolving parameter set of the new design, we can map that to a previous battery design to see how the patterns compare. And we also leverage the oracle because it is trained using previous designs. Also, the interpreter can tell us about the parameters of interest to us. For example, if we think there are 11 or 14 parameters that are very important to the battery lifetime, we use the interpreter to get those parameters from our initial testing data.

Then we leverage the oracle to take these parameter sets as inputs and predict the lifetime of our design. But of course, the oracle cannot give us precise results in the first round, so we have to iterate to reduce the uncertainty and get better results over time.

Since we know that’s not too accurate, we do a second run from the learner. I get the results and you tell me the lifetime of the three battery designs I have selected. But I'm still uncertain about certain designs or testing conditions. So, I might pick two more cells and redo the procedure once again — that is how people learn. We iterate for different conditions.

Tech Briefs: Once you come up with your prediction, how do you prove that you're right?

Song: So, in the first round for new batteries, we choose the representative testing conditions intuitively. For example, we care about low temperature performance, high temperature performance, medium temperature performance, high discharge rate, and low discharge rate. In the second round, we have some quantitative data to help us choose further cells. We then use Gaussian process regression to give us the predictions’ uncertainty. Then we choose the most uncertain predictions and run those conditions again to confirm the original predictions.

This is common practice in the field, so most people trust it. However, we received a comment when we published our paper asking how we can confirm that our uncertainty quantification is correct? There are different methods, to be honest, and although we didn't try all of them, we plan to do so in the future.

Tech Briefs: Suppose a battery manufacturer is interested in your results. Will he trust you enough to start building batteries on an assembly line based on your prediction?

Song: That's a good question. I think so, at least for one battery company that we have been collaborating with for almost four years. And in our collaboration, they provide all the data to us. We have been discussing the results with their engineers, and they think the results are trustworthy. So, if they have a new battery design, especially for one that is not very different from a previous design, our framework can work.

But I also want to point out the limitations of our study, which is very important. For example, we cannot positively specify the limitations if the new battery design is very different from previous ones. The framework is not magic — if the domain knowledge you have now is very different from a new design, how can you make a prediction?

We haven't obtained a methodology to quantify these questions — it’s out of the scope of this work. But I think it's worth going further. For example, our study only covers traditional lithium batteries and does not include some of the more advanced technologies like solid-state or lithium-metal.

Tech Briefs: So, you're just covering standard lithium-ion batteries?

Song: Exactly. Currently, most products use standard lithium-ion batteries. But even for commonly used lithium battery cells, when you have new designs, an electric vehicle company or battery company will do one-year testing for each design. So, our tool is very useful right now. But the limitation is what to do when we have very different designs.

Tech Briefs: You’ve also mentioned physics-based analysis. What does that mean?

Song: It means we make use of a physics-based model — a so-called first-principle model. For example, some very basic physical equations like the Fick's law diffusion equation. Using these, we can generate features that are physically interpretable. For example, in many previous papers, they just use features derived from voltage and current measurements, but they don't know the specific physical meaning of those features.

By incorporating physical modeling along with the parameter calibration, based on the parameter following a certain trend, I can tell you either why your battery cell is durable enough, or that you may have some type of degradation mechanism. It will make the prediction more trustworthy and more explainable.

If you just use machine learning, and you don't pay attention to physics, the dataset you train your machine learning model on might give you a prediction that can work for 95 percent of the cases. But that means, although maybe it works for you, it might not work for me. And if it doesn't work for me, it may give me a ridiculous prediction, inconsistent with the physics, and that’s not tolerable in practice. Regarding accuracy, probably a one percent error is fine, maybe even a two percent error. I just want to give you reliable estimates, not ridiculous, crazy estimates.

Tech Briefs: How do you input the physical parameters into your system?

Song: We have a model indicating the system response. I know that the parameters have a relationship with the response. So, we use our interpreter to get a statistical distribution of those parameters, because it’s very hard to calibrate such a large number of parameters precisely. For an existing good design and dataset, you just do the parameter calibration, the distribution, and you try to map your parameter distribution and trends to the lifetime of your battery cells. Because for the existing battery design, since you have the whole lifecycle test data, you can see the actual relationships. Then we try to leverage this kind of relationship for our new battery designs.

Tech Briefs: What are your next steps?

Song: My plan regarding this work is to figure out the limitations of the framework. We can only test limited battery designs, not every single one. Right now, we are talking about solid state batteries, lithium metal, lithium sulfur batteries, and others, and they all have different chemistries. So, we want to see whether this framework can work well for new kinds of batteries. If not, we want to explore what we need to do. I think the framework will still work well, but we will have to make changes because different battery chemistries will have very different internal physics.

For battery people, the most important component is the interpreter. We have a lot of things to do in this field. For example, we can enhance our physics-based modeling to enhance its performance for modeling solid-state batteries.