Enlarge /. The seemingly simple charging is becoming more and more complex.
Batteries usually involve a lot of compromises. They can have a high capacity, but this means more weight and slower loading. Or you can charge quickly and see how the life of your battery decreases with each cycle. There are ways to optimize performance – to find out how quickly charging can be done without reducing battery life – but this varies from product to product and requires extensive testing to identify it.
But maybe testing is not as extensive thanks to a new system that was described in Nature magazine. The system uses a combination of machine learning and Bayesian inference to quickly determine the optimal charge pattern for each battery and significantly reduce testing effort.
Not so fast
Fast charging is obviously useful for everything from the phone to the car. However, if a battery is charged quickly, it doesn't store its ions quite as efficiently. The total capacity will decrease and there is a possibility of permanent damage as part of the lithium fails and is no longer available for future use.
However, there are ways to change the loading profile to avoid this problem. For example, it might be possible to start charging slowly and create an orderly lithium storage and then switch to fast charging that builds on it before the charging rate is slowed down again to efficiently pack the last bit of lithium. Modern chargers have enough computing power to manage a charging process that optimizes the speed compared to the battery performance. The performance of all batteries decreases over time, but the right profile minimizes them.
The problem is finding the right charging profile. At the moment, we just have to do empirical tests: run a series of batteries through many charge / discharge cycles and monitor how their performance changes over time. Because there are many potential charging profiles and the performance degradation is gradual, hundreds of batteries need to be put through enough charge / discharge cycles to get them close to the end of their lives. To make matters worse, the profile is different for each battery type. So if you know what kind of charging works well for your phone, you won't necessarily learn how to charge a phone from another manufacturer.
The new work, which was done in a large collaboration, was an attempt to reduce the time to test a particular battery.
The setup the researchers use includes standard battery test hardware that allows them to send multiple batteries simultaneously through repeated charge / discharge cycles. In addition, most of the action takes place in software.
An important software component is known as a Bayesian Optimizer or BO. The BO balances two competing interests: in order to find the best loading profile, as many profiles as possible have to be tested, and the best profile is likely to be close to a profile that you have already identified as good. If you mistreat this balance, you will end up exploring the whole area for a decent solution, but missing out on a number of better solutions elsewhere in the load profile row.
Bayesian statistics take into account previous information so that it can use the knowledge from the first test rounds to ensure that both future rounds examine more solutions at the same time and additional tests focus on the best solutions from previous rounds.
A Bayesian optimizer alone would simply increase the efficiency with which a number of loading profiles are tested – good, but not particularly exciting. In this case, however, the researchers paired it with a machine learning algorithm that uses and uses the voltage profile during discharges to predict the future life of the battery. In previous work, this algorithm was able to successfully predict lifetime performance with only 100 data cycles. As a result, testing of a battery pack is reduced from 40 days to 16 days.
This is good for a single test round. However, keep in mind that the goal is to both examine most cargo profiles and test all profiles related to the successful solutions of the first round. Running just a few rounds of this type of testing could take almost half a year to determine the best charging profile. And after six months, most companies prepare to work on a new product design – often with a completely different battery.
Tests in practice
To show that the system actually worked, the research team used a test device with 48 batteries and tested a set of 224 quick-charge profiles that performed a 17-minute charge. This usually shortens the life of the battery considerably. After just two rounds of testing with 100 cycles, the researchers were able to understand the general outline of the best solutions and examine most of the potential profiles under consideration.
In this case, it turned out that the best solutions were linear charging profiles, in which the charging rate was kept constant throughout the cycle. However, as mentioned earlier, this is likely to be different if a different battery is used. And even a single type of battery such as lithium-ion can differ dramatically in terms of its physical structure, the electrolyte used, the electrode chemistry, etc. After all, there are clearly applications in which different charging profiles would be prioritized. An electric car may need to be charged quickly during transportation. However, if it is parked at home, it may be better if the profile optimizes battery life. There is no reason why this test setup cannot handle both.
One of the most striking things about it is that even if all of this optimization work is done, it is completely invisible to most users. While users may find that their device charges faster than usual, they know nothing about the electronics in their charging hardware that change the charging profile while getting feedback on battery status.
Nature, 2020. DOI: 10.1038 / s41586-020-1994-5 (About DOIs).