TU-TESL

• 22

At present, enterprises and scientists are actively developing new batteries or looking for new technologies to optimize battery manufacturing. Foreign media reported that Stanford University and researchers have developed a new machine learning method, which can accelerate the development of electric vehicle batteries. Specifically, the research teams at MIT Stanford and Toyota Research Institute have developed a machine learning based approach to reduce the battery charging test time from nearly two years to 16 days, which is nearly 15 times shorter, helping to accelerate the development of new batteries. At each stage of the battery development process, new technologies must be tested for months or even years to determine how long they will last. Designing ultra fast charging batteries is a major challenge, mainly because it is difficult to keep them in use. Faster charging intensity will put more pressure on the battery, which usually leads to premature failure of the battery. To this end, MIT and the Toyota Institute hope to find the best way to charge the EV battery within 10 minutes, so as to maximize the overall service life of the battery. To find the best way, the team used AI to help classify various charging tests. The team published the study Tuesday in the journal Nature, which shows how patented AI programs predict how batteries react differently to charging methods. From the beginning, the team found that fast charging optimization required multiple trial and error tests - inefficient for humans, but perfect for machines. First, the battery is tested. The first 100 cycles of cycle data (especially electrochemical measurements such as voltage and capacitance) are used as inputs for the prediction of early cycle life results. These cycle life predictions from the machine learning (ML) model are then sent to the Bo algorithm, which proposes to test the next protocol with higher estimated life by balancing the competitive requirements of exploration (test protocol with high uncertainty of estimated life) and development (test protocol). Repeat this process until the test budget runs out. In this method, early prediction reduces the number of cycles required for each test cell, while the optimal experimental design reduces the number of experiments required. The small training data set recycled to the failed battery can be used not only to train the early result predictor, but also to set the Bo super parameters. In the future, the design of battery material and process can also be integrated into the closed-loop system. Because the machine learning system can find the pattern of predicting the battery's sustainable use time in the early data after receiving the battery's training of several failed cycles. Machine learning reduces the number of methods they have to test. Computers do not test every possible charging method equally, nor do they rely on intuition to test, but learn from their experience that they can quickly find the best test protocol. "When we talk to material scientists and people who work in batteries, we realize that no one actually uses more complex AI in this field, so we think it's promising," said ermon, a computer science professor at Stanford University "You can apply different voltages, different currents, different intensities - they may all charge the battery at the same time, but some may damage the internal components of the battery," he said The method is expected to speed up every aspect of battery development, from designing the chemical properties of batteries to determining their size and shape to finding better manufacturing and storage systems, the researchers said. "We've come up with a way to dramatically speed up the test process for ultra fast charging, which can be applied to many other problems that may currently hinder battery development for months or years." Peter Atia, CO head of the project, said. "This is a new way to develop batteries," said Patrick herring, a scientist at Toyota Research Institute. Having data that can be shared and analyzed automatically between many people in academia and industry can speed innovation. " He added that optimizing the rest of the battery development process through machine learning, battery development and the emergence of newer and better technologies would accelerate by an order of magnitude or more. The project was supported by Stanford University, the Toyota Institute, the National Science Foundation, the U.S. Department of energy and Microsoft.