Using Statistics and Machine Learning to Test the Durability of Energy Storage Solutions

There is a lot of hype these days in the media about artificial intelligence (AI) and machine learning (ML), but at Berkeley Lab, AI and ML have played an integral role in research projects for many years.

One current example of a research project with an important AI/ML component is a collaboration between the Energy Technologies and Computing Sciences Areas. Steve Harris, Principal Scientific Engineering Associate with ETA’s Energy Storage and Distributed Resources Division, and Marcus Noack, a research scientist with the CSA’s Applied Mathematics and Computational Research Division, are working together to develop a unique AI/ML-driven solution for the rapid assessment of the durability of novel energy storage solutions.

Left to right: Steve Harris and Marcus Noack

Utilities are seeking new energy storage technologies to back up wind and solar power, which are rapidly taking up a greater share of energy production in many states. Many new solutions are being developed to ensure adequate long-duration energy storage (LDES) for wind and solar energy: thermal storage, pumping air into caves, and spin gyroscopes, just to name a few. These energy storage solutions need to be warranted for 15 or more years in order to make the investment worthwhile. But since these technologies have only been around for a few years, or have just been introduced to the market, utilities need a way to understand the durability of these solutions without waiting multiple years.

Steve, who joined the Lab in 2013 after a career at GM and Ford, first saw the potential of AI/ML to address this problem four years ago when he was working on a project with Stanford University. The project collected massive amounts of battery data which MIT’s statistics department analyzed using machine learning techniques. Following that project, he thought of applying machine learning to the energy storage durability assessment project he had been considering. In particular, he wanted to offer industry information that current solutions, which only offer information about average lifespans, did not: he wanted to offer information that can help with (1) setting a warranty, whose cost depends almost entirely on the fraction of the devices that fail well before the expected life, and (2) estimating the value of second use, which depends almost entirely on performance long after the expected life. In short, to assess whether a technology was viable in the marketplace, he needed to be able to provide information on the entire distribution of the technology’s durability.

Steve Meets Marcus, and an AI/ML Research Project is Born

Steve’s opportunity to move forward on this idea came in 2021 when Noel Bakhtian, then leader of the Energy Storage Center, organized get-to-know-you meetings across divisions. Steve presented his idea, and Marcus stood up and said, “I can do that.”

A research scientist with the Computing Sciences Area’s Applied Mathematics and Computational Research Division, Marcus didn’t have a battery background, but he did have AI/ML and statistical mathematics expertise. Hired six years ago into the Center for Advanced Mathematics for Energy Research Applications (CAMERA), his research focuses on estimating probability distributions, which is a growing field in AI/ML, as well as on developing techniques for optimal and autonomous data acquisition.

Together, Steve and Marcus knew they could tackle the challenge with their complementary skills. They planned to start with the assessment of batteries, where the data is more plentiful, and then use the model for long-duration energy storage. (In fact, the model Steve and Marcus are building is transferable to other projects that need analysis of sets of cycle data where it is helpful to have information about the distribution of a technology’s lifespan). They planned to apply statistics and ML, in the form of stochastic modeling, to published lithium battery durability data to estimate lifespan distributions, while minimizing the necessary time and resources to do so.

Said Marcus, “Every math problem starts with looking at sets. Once you understand the space you’re dealing with, you then think about functional rules in that space. Once you understand both, you can decide what math techniques you want to use to represent the functional relationships in this space. In this case, we looked at the capacity of batteries, how they are being used, and how they are built, in order to understand their durability. You can create a model that you can query for other situations.”

Steve and Marcus applied for LDRD funding in 2022, with Steve acting as principal investigator, and in May got the good news that their project was funded. They kicked off their project last October and have already published a first paper, with a second planned in the next few months.

How Researchers Manage the Risks of AI/ML

What about the risks that have been ascribed to generative AI/ML solutions like chatGPT, including the problems of “hallucinations” reported in the media? Are these a concern for researchers considering employing AI/ML in their work?

“AI/ML in research is very different from solutions like chatGPT,” said Marcus. “With chatGPT, you are harvesting information that you don’t have control over. Research scientists have control of the data that we put in; we know what data has been collected and how it has been collected.”

Another risk is what he calls “embedding risk” – essentially, the risk of assigning similarities to things that are not similar. For instance, with a certain set of weather data, it may be reasonable to predict what the weather will be like two years from now but not ten years from now. Systems like chatGPT may not be able to assess this risk adequately.

“Researchers have certain access points so that they can check what the AI/ML system is learning, how it’s learning, and if it is learning correctly. The data risk is smaller, and the embedding risk can be managed,” said Marcus. “Researchers can manage the data and embedding risks while applying the very significant capabilities of AI/ML to their work.”

According to Steve, the role that AI/ML will play in science cannot be overstated. Said Steve, “AI/ML is changing all of science. These days, a proposal that includes an AI/ML component has a greater chance of getting funded than one that doesn’t.”

In fact, he tells the grad students and the postdocs that he mentors: “Learn enough ML so you can go out and do it. It is the biggest change to science since the introduction of computers. Not knowing how to apply ML will be like not knowing how to code today.”

Additional stories and links about AI/ML at the Lab:

Sitemap