Batteries tend to involve lots of trade-offs. You can have high capacity, but it means more weight and a slower charge. Or you can charge quickly and see the lifetime of your battery drop with each cycle. There are ways to optimize performance—figuring out the fastest charging you can do without cutting into the battery life—but that varies from product to product and requires extensive testing to identify.
But perhaps that testing is not so extensive, thanks to a new system described in the journal Nature. The system uses a combination of machine learning and Bayesian inference to rapidly zero in on the optimal charging pattern for any battery, cutting the amount of testing needed down considerably.
Not so fast
Fast charging is obviously useful for everything from phones to cars. But when a battery is subjected to fast charging, it doesn't store its ions quite as efficiently. The overall capacity will go down, and there's the potential for permanent damage, as some of the lithium ends up precipitating out and becoming unavailable for future use.
There are, however, ways of altering the charging profile to avoid this issue. For example, it might be possible to start charging slowly and generate some ordered lithium storage and then switch to rapid charging that builds on these before slowing the charge rate again to pack the last bit of lithium in efficiently. Modern chargers have enough processing power to manage a charging process that's designed to optimize speed against battery performance. All batteries see performance drop over time, but the right profile will minimize it.
The problem is identifying the right charging profile. At the moment, the only way we have to find it is to do empirical tests: run a bunch of batteries through a lot of charge/discharge cycles and monitor how their performance changes over time. Since there are a lot of potential charging profiles, and the performance decay is gradual, the process ends up requiring that hundreds of batteries have to be sent through enough charge/discharge cycles to take them to near their end-of-life point. Making matters worse, the profile will be different for each battery type, so learning what sort of charging works well for your cell phone won't necessarily tell us how to charge a phone from a different manufacturer.
The new work, done by a large collaboration, was an attempt to cut down on the time involved in testing a given battery.
Learning Bayesians
The setup the researchers use does involve standard battery-testing hardware, allowing them to send multiple batteries through repeated charge/discharge cycles at the same time. But beyond that, most of the action takes place in software.
One key software component is called a Bayesian Optimizer, or BO. The BO balances two competing interests: finding the best charging profile will mean testing as many profiles as possible, and the best profile is likely to be somewhere near one you've already identified as being good. Handle this balance poorly and you'll end up exploring all the area around a decent solution but miss a cluster of better solutions elsewhere in the set of charging profiles.
Bayesian statistics is designed to take prior information into account so it can use knowledge gained from the first few rounds of testing to ensure that both future rounds simultaneously explore more solutions while focusing additional tests near the best solutions from earlier rounds.
On its own, a Bayesian optimizer would simply increase the efficiency with which a set of charging profiles is tested—good, but not especially exciting. But in this case, the researchers coupled it with a machine-learning algorithm that takes the voltage profile seen during discharges and uses that to predict the future lifetime of the battery. In previous work, this algorithm was able to successfully predict lifetime performance using just 100 cycles of data. This has the effect of cutting testing of a set of batteries from 40 days down to 16.
That's good for a single round of testing. But remember that the goal is to both explore most of the set of charge profiles and to test all of the profiles around the successful solutions found in the first round. Doing just a few rounds of that sort of testing could mean nearly a half-year spent identifying the best charging profile. And by the time six months have passed, most companies are gearing up to work on a new product design—often one involving a different battery entirely.
Real-world testing
To show that the system actually works, the research team used a 48-battery testing device and tested a set of 224 fast-charging profiles that performed a 17-minute charge. This typically shortens the lifetime of the battery dramatically. After just two rounds of testing using 100 cycles, the researchers were able to understand the general outlines of the best solutions and had explored most of the potential profiles under consideration.
As it turns out, in this case, the best solutions were linear charging profiles, where the rate of charge was kept constant throughout the cycle. As mentioned earlier, however, that will likely be different if a different battery is used. And even a single type of battery like lithium-ion can differ dramatically in terms of its physical structure, the electrolyte used, the electrode chemistries, and so on. Finally, there are clearly applications in which different charging profiles would end up being prioritized. An electric car might need fast charging while in transit, but when parked at home, it might do better with a profile that optimized battery lifetime. There's no reason this test setup couldn't handle both.
One of the more striking things about this is that, even if all this optimization work is done, it will end up being completely invisible to most users. While users might notice that their device charges faster than they're used to, they won't know anything about the electronics in their charging hardware that alters the charging profile while receiving feedback on the battery's status.
Nature, 2020. DOI: 10.1038/s41586-020-1994-5 (About DOIs).
reader comments
37