Argonne's Machine Learning Algorithm Can Tell You How Long a Battery Will Last — From a Single Cycle

Hitting a mean absolute error of 103 cycles after just one full charge-discharge cycle, this ML system could boost battery development.

Researchers at the Argonne National Laboratory have developed a machine learning system capable of accurately predicting battery lifetimes — something they claim could help to reduce the cost of battery development.

"For every different kind of battery application, from cell phones to electric vehicles to grid storage, battery lifetime is of fundamental importance for every consumer," explains computational scientist Noah Paulson, co-author, of the study's focus. "Having to cycle a battery thousands of times until it fails can take years; our method creates a kind of computational test kitchen where we can quickly establish how different batteries are going to perform."

Designed to bypass the need to actually perform full charge-discharge cycles in order to test a given battery, the team's machine learning models proved capable of accurately estimating the lifespan of a lithium-ion batteries using a variety of cathode chemistries and numerous electrolyte and anode compositions from as few as 100 charge-discharge cycles — and, impressively, could get to within a mean absolute error (MAE) of 103 cycles when predicting lifespan based on only a single charge-discharge cycle.

"The reality is that batteries don’t last forever, and how long they last depends on the way that we use them, as well as their design and their chemistry," Paulson explains. "Until now, there’s really not been a great way to know how long a battery is going to last. People are going to want to know how long they have until they have to spend money on a new battery.

"We had batteries that represented different chemistries, that have different ways that they would degrade and fail. The value of this study is that it gave us signals that are characteristic of how different batteries perform."

Those signals go beyond simple prediction of known battery chemistries, too. The team's algorithm can be trained on known chemistries then predict the performance of unknown chemistries, something Paulson and colleagues claim will help to lower the cost of and speed up battery development by pointing researchers in the direction of lifetime-enhancing chemical combinations not yet tried.

"Say you have a new material, and you cycle it a few times. You could use our algorithm to predict its longevity, and then make decisions as to whether you want to continue to cycle it experimentally or not," Paulson explains. "If you’re a researcher in a lab, you can discover and test many more materials in a shorter time because you have a faster way to evaluate them," adds co-author Susan Babinec.

The team's work has been published under closed-access terms in the Journal of Power Sources.

Gareth Halfacree
Freelance journalist, technical author, hacker, tinkerer, erstwhile sysadmin. For hire: freelance@halfacree.co.uk.
Latest articles
Sponsored articles
Related articles
Latest articles
Read more
Related articles