INSIDE A LAB at Stanford University’s Precourt Institute for Energy, there are a half dozen refrigerator-sized cabinets designed to kill batteries as fast as they can. Each holds around 100 lithium-ion cells secured in trays that can charge and discharge the batteries dozens of times per day. Ordinarily, the batteries that go into these electrochemical torture chambers would be found inside gadgets or electric vehicles, but when they’re put in these hulking machines, they aren’t powering anything at all. Instead, energy is dumped in and out of these cells as fast as possible to generate reams of performance data that will teach artificial intelligence how to build a better battery.
In 2019, a team of researchers from Stanford, MIT, and the Toyota Research Institute used AI trained on data generated from these machines to predict the performance of lithium-ion batteries over the lifetime of the cells before their performance had started to slip. Ordinarily, AI would need data from after a battery had started to degrade in order to predict how it would perform in the future. It might take months to cycle the battery enough times to get that data. But the researchers’ AI could predict lifetime performance after only hours of data collection, while the battery was still at its peak. “Prior to our work, nobody thought that was possible,” says William Chueh, a materials scientist at Stanford and one of the lead authors of the 2019 paper. And earlier this year, Chueh and his colleagues did it again. In a paper published in Nature in February, Chueh and his colleagues described an experiment in which an AI was able to discover the optimal method for 10-minute fast-charging a lithium-ion battery.
Many experts think fast-charging batteries will be critical for electric vehicle adoption, but dumping enough energy to recharge a cell in the same amount of time it takes to fill up a tank of gas can quickly kill its performance. To get fast-charging batteries out of the lab and into the real world means finding the sweet spot between charge speed and battery lifetime. The problem is that there is effectively an infinite number of ways to deliver charge to a battery; Chueh compares it to searching for the best way to pour water into a bucket. Experimentally sifting through all those possibilities to find the best one is a slow and arduous task—but that’s where AI can help.
In their research, Chueh and his colleagues managed to optimize a fast-charging protocol for a lithium-ion battery in less than a month; to achieve those same results without the aid of AI would usually take around two years. “At the end of the day, we see our job as accelerating the pace of battery R&D,” says Chueh. “Whether it’s discovering new chemistry or finding a way to make a safer battery, it’s all very time consuming. We’re trying to save time.”
Over the past decade or so, the performance of batteries has skyrocketed and their cost has plummeted. Given that many experts see the electrification of everything as key to decarbonizing our energy systems, this is good news. But for researchers like Chueh, the pace of battery innovation isn’t happening fast enough. The reason is simple: batteries are extremely complex. To build a better battery means ruthlessly optimizing at every step in the production process. It’s all about using less expensive raw materials, better chemistry, more efficient manufacturing techniques. But there are a lot of parameters that can be optimized. And often an improvement in one area—say, energy density—will come at a cost of making gains in another area, like charge rate.
Finding optimal solutions in a huge search space is exactly the type of problem AI was built to solve. But until recently, battery-building AIs were hampered by a lack of data. “Historically, battery data has been very difficult to acquire because it’s not shared between researchers and companies,” says Bruis van Vlijmen, a data scientist working on battery analytics at Stanford. “There’s a high level of secrecy or proprietary information.” Following their 2019 paper, Chueh and his colleagues made all of their battery data publicly available so it could be used by other researchers to train their own AI algorithms. At the time, it was the largest collection of battery performance data ever released.
For Ian Foster, the director of the data science and learning division at Argonne National Laboratory, the lack of quality data is a familiar problem. For the past few years, Foster and his colleagues at the lab have been building a database of molecules that can be prowled by a machine learning algorithm to hunt for chemicals that might lead to improved performance in a battery’s electrolyte, the stuff that sits between the electrodes. Like the other elements in a battery, electrolyte chemistries can be tweaked to boost desirable properties like energy density or reduce undesirable ones like its toxicity. “Historically, identifying new electrolyte materials has been very much a trial and error process,” says Foster. “Our goal is to apply AI methods to explore the essentially infinite space of possible materials.”
In late 2019, the Argonne team published a pair of papers that detailed how they used an existing database of 133,000 organic molecules and the lab’s supercomputer to create ultraprecise simulations of the properties of these molecules that have up to nine “heavy,” or non-hydrogen, atoms. Their idea was to use this database to train a machine learning algorithm to find molecules with desirable properties in a relatively small dataset so it could explore a much larger database of potential materials. The molecules in most battery electrolytes may have upward of 20 heavy atoms, and there are a lot of ways those atoms can be combined. For example, another database of organic molecules that have up to 17 heavy atoms consists of 166 billion candidates. That would be an unreasonably large space for an AI to seek out promising candidates without having a good idea of what it was looking for.
Foster says it’s still early days for Argonne’s electrolyte hunting algorithm. It hasn’t identified any new materials just yet, but when it does the next step will be to create a physical cell using that electrolyte material for experimentation. The data from those experiments can then be used to further refine the algorithm and help it narrow its search to still better candidates. “The process of actually going from a very large number of possible electrolytes to one that actually will be deployed in millions of cars is a long one,” says Foster. “The goal of machine learning is to accelerate the experimentation process.”
In the meantime, Foster’s team is working with battery scientists at a dozen research institutions and companies to facilitate sharing stats across organizations. The group hopes to use a platform developed at the University of Chicago called Data Station that allows researchers to train machine learning models on a pool of information contributed by different groups without ever giving outsiders direct access to their data. A machine learning model is uploaded to the platform, trained on the data, and then returns to the researchers. Those scientists don’t know the specifics of the data, but they can tell whether exposure to that data improved the model’s ability to make predictions about batteries. Foster and his collaborators hope this will assuage people’s fears about losing proprietary data to competitors while still allowing the creation of the massive data sets.
But even without enormous shared databases, the use of AI in battery development is already heating up. As detailed in a paper published this summer in Frontiers in Energy Research, just in the past year AI has been used for a staggering number of applications in battery research. On the materials side, it’s been used to study molecules that can stabilize lithium metal anodes, which can drastically boost energy density but currently come with a lot of safety concerns. Machine learning was also used to discover potential cathode coatings to improve the performance of batteries with solid electrolytes, which are safer that the liquid electrolytes found in batteries today. AI has also been used to improve researchers’ understanding of existing batteries by optimizing battery management systems and creating precise mathematical models of batteries to simulate their performance in EVs. An AI even wrote a book summarizing current research on lithium-ion batteries.
“There is a lot of untapped potential in existing battery materials, which we can harness by using better software to ‘program’ the battery,” says Alpha Lee, a statistical physicist at the University of Cambridge whose recent research has used machine learning to find new predictors of a battery’s health. “Innovations in battery software will benefit from the degree of scalability that we saw in the digital revolution and usher in a new era of energy storage technologies.”
The next step is to take these machine learning methods out of the laboratory and use them to make batteries that will power our gadgets and cars. InoBat, a Slovakia-based company founded in 2018, may be leading the way. The company is using an AI-powered research platform developed by California-based Wildcat Discovery Technologies to rapidly prototype new battery chemistries to make bespoke cells for electric vehicles. According to InoBat’s CEO Marian Bocek, the AI platform allows for a holistic exploration of new lithium-ion chemistries, which has the potential to dramatically speed up the discovery process. In other words, rather than tweaking one battery component at a time and exhaustively testing each iteration, the AI can simulate a battery’s performance when several different variables have been modified at once.
“The road to discovery of new cell chemistry is 10 times faster compared to a traditional lab,” says Bocek, who compares InoBat’s AI-fueled research to the use of automated drug discovery in the pharmaceutical industry. “We’re moving away from the ‘one size fits all’ model that is dominating the EV industry.”
InoBat unveiled its first “intelligent battery” designed with AI last week. During the announcement, Bocek claimed that the battery could boost the range of a “best-in-class” EV by nearly 20 percent. But don’t expect to find it in the battery pack of an average EV anytime soon. Unlike major producers of lithium-ion cells, such as Panasonic or Samsung, InoBat is more of a battery boutique. The company focuses on specialized vehicles such as high-performance EVs or electric aircraft, and can do low-volume production to develop cells that meet a customer’s specific needs. “We are the only player like this in the market that has the capacity to develop a customized solution in terms of cell format and energy density,” says Bocek.
Bocek says the company’s first pilot plant will start cranking out batteries by the end of next year. Initially, the plant will produce just 100 megawatt-hours of AI-designed batteries per year. To put that in perspective, that’s about one half of 1 percent of the production volume of Tesla’s Gigafactory in Nevada. Bocek says the company has plans to scale its production up to a 10 gigawatt-hour facility within five years. That will put it on par with planned output at Tesla’s new pilot plant in California that company officials announced at the Battery Day event last month.
Unleashing AI on battery development is good news for a warming world. Battery storage is a key factor in increasing the amount of renewable energy on the grid, and when it comes to decarbonizing our energy supply, time is of the essence. After decades of plodding progress, AI-driven battery research promises to finally pick up the pace. “This is all tied back to decarbonisation,” says Chueh. “We want to get there quickly because we don’t have much time left.”
