AI predicts battery life

Reading time: 3 minutes


Using machine learning to sort through hundreds of millions of data points, US researchers came up with models to predict useful battery life.

It’s an annoyance many consumers have had to deal with: one day your phone battery easily lasts a whole day, seemingly the next it runs out of juice even before you started your evening commute. Clearly, the capacity of lithium-ion batteries can suddenly take a sharp turn downward. How many cycles a battery lasts tends to vary widely though. Sometimes you get lucky when buying a new phone, sometimes you don’t.

Manufacturers of cell phone batteries would love to be able to tell which ones will last, say, over two years and which ones won’t. They could sell the best ones to premium brands, send the rest to makers of budget devices or use them for less demanding applications altogether.

Currently, there is no way to do this, however. Battery testing involves charging and discharging a battery until they fail. After this time-consuming process, the battery is useless, while it still may not be clear why it lasted as long as it did. Apart from battery manufacturers, this hurts the development of battery technology in general. “Testing is an expensive bottleneck in battery research,” says Stanford researcher Peter Attia.

Attia and co-workers from Stanford, MIT and the Toyota Research Institute developed a method to predict battery lives without waning their capacity completely. Using machine learning to analyze hundreds of millions of data points, they came up with models that can reliably predict how many more cycles any particular lithium-ion battery will last.

Time consuming

Generally, two mechanisms contribute to performance degradation: a progressive loss of storage capacity and a progressive increase of impedance, which causes battery power to decline. Both processes are inextricably linked, however. “Lithium-ion batteries are complex systems, which means that capacity fade and impedance increase arise from several interacting processes. Most of these processes cannot be studied independently and occur at similar timescales, which complicates the investigation of aging mechanisms. Battery health, therefore, cannot be determined from a single direct measurement,” explains Maitane Berecibar from the Vrije Universiteit in a Nature commentary.

Hence the invocation of the power of numbers. 124 commercial lithium-ion batteries were charged and discharged until they reached the end of their useful life, which the researchers had defined as a capacity loss of 20 percent. The batteries were charged in a variety of different ways in order to ensure different types of degradation behavior were included in the data set. The discharge conditions were identical.

In the first thousand charge-discharge cycles, a significant number of batteries start losing capacity. Graph courtesy of Nature Energy

One of the resulting models accurately predicted the number of cycles left for about 90 percent of the batteries. This prediction was based on the data for the first hundred charge-discharge cycles, which is well before battery performance starts to degrade. The best batteries in the tests lasted 2300 cycles.

Another model used data from only the first five cycles to classify a battery as either having a short (ie less than 550 cycles) or a long lifetime (ie more than 550 cycles). In this case, the model’s predictions were correct about 95 percent of the time.

Apart from matching individual batteries with applications according to their performance, the models could be used for shortening the time needed to validate new battery designs. Testing is currently the most time-consuming part of that process. Yet another possibility is optimizing battery manufacturing. “The last step in manufacturing batteries is called ‘formation’, which can take days to weeks,” Attia says. “Using our approach could shorten that significantly and lower the production cost.”