A.I. Method Could 'Supercharge' Electric Vehicle Battery Development
Advances in electric vehicle batteries have often been limited by the enormous bottleneck of evaluation times. At each stage of the battery development process, new technologies must be tested for months or up to years to determine how long they will last. Now, Stanford University researchers have developed a machine learning-based method that cuts these testing times by 98 percent. The researchers tested the method on battery charge speed, but say that it can be applied to many other parts of battery development. The breakthrough method could make the dreams of recharging an electric battery in the time it takes to stop at a gas station a more likely reality.
Transcript
00:00:06 what if you could recharge your electric car in under 10 minutes about the time takes to stop at gas station the overall goal of this research is to develop methods that allow us to accelerate battery testing but without degrading it a large amount when batteries are charged at a very fast rate it puts greatest strain on the battery which causes the degradation on their
00:00:34 overall life the main bottleneck with advances in electric vehicle batteries is testing times going from the laboratory to an actual deployment in evey can take years the method we've developed is a machine learning technique to reduce testing times by about 98% so typically it takes about a couple of months to finish an experiment in this case and we were able to reduce
00:00:56 that time down to four days we've done this I think through a variety of different approaches one is to accelerate the test itself and then the other is to accelerate the entire process of generating those tests and determining what tests to run so instead of testing hundreds or possibly millions of combinations we test just a few and use a machine
00:01:17 learning algorithm to determine what the outcomes would be on all the others the method is very general and it can be applied to many other cases that involve the time-consuming experiments so there are basically two components to the research it's the researcher through the program that based only on a few cycles of charging we'd figure out what's the
00:01:36 lifetime of the battery and the other component of the research was an algorithm which figures out which charging protocols to test and when so this is an exciting an academic and industrial collaboration between Stanford MIT and the Toyota Research Institute the Stanford groups were focused first on experiments providing the data at your eye
00:01:58 they've helped with the design of the data collection pipeline and the cloud infrastructure the group at MIT was mostly involved in making early predictions the current work improves the efficiency of the optimization the exact details depend on the optimization problem but broadly speaking we are reducing the number of experiments to the bare minimum needed
00:02:16 to reach a conclusion the specific goal of this research was to optimize fast-charging can we charge a battery in ten minutes and what's the best way to do that the data that is being processed for these machine learning algorithms is being packaged into cycles we take each time we charge and discharge the battery is a distinct unit and we both clean and
00:02:38 standardize that data there are substantial variations in performance from one battery to the next which makes it difficult to apply traditional optimization methods however there are machine learning algorithms that are specifically designed to address problems of this nature it also helped us identify novel charging protocols which we did not know existed that could
00:03:02 help us charge a battery much faster using a data pipeline is the best practice for data at scale especially as scientific experiments start generating more and more data the volume increases this becomes a much more important practice I think one of the reasons this was not done before is that it really required a lot of interdisciplinary collaboration between and expertise both
00:03:24 in the machine learning world and what kind of algorithms we should use for this particular application and a lot of expertise in terms of like battery technology this provides a better way to do science and experiments in the future as science becomes much more data intensive this provides the ability to utilize that data in an optimal manner how come we use AI and machine learning
00:03:48 optimal experimental design and kind of technologies that we've just talked about not only to optimize for something but actually help us come up with a better understanding of the physical processes the next steps for this project are to apply this method to much more challenging problems problems in manufacturing of batteries and problems in real-world use cases like when do you
00:04:09 charge your battery or how do you keep it charged we learned a lot through this process both about the application of these algorithms and also that the algorithm found that delivering the highest current in the middle of the charge actually results in batteries with the longest lifetime [Music]