创新背景

如果手机电池制造商能够分辨出哪些电池可以使用至少两年，那么他们就可以只把那些电池卖给手机制造商，把其余的卖给要求不那么高的设备制造商。

创新过程

斯坦福大学、麻省理工学院和丰田研究所的科学家们发现，结合综合实验数据和人工智能，揭示了在锂离子电池容量开始衰退之前准确预测其使用寿命的关键。在研究人员用数亿个电池充电和放电的数据点训练他们的机器学习模型后，算法根据电压下降和早期循环中的其他一些因素，预测每个电池还能持续多少次循环。

这些预测与细胞实际存活的周期数相差不到9%。此外，该算法仅根据前5次充放电周期将电池分为长寿命或短寿命两类。在这里，预测的准确率达到了95%。

《自然能源》杂志上的这种机器学习方法可以加快新电池设计的研发，并减少生产时间和成本，以及其他应用。研究人员已经将这一规模最大的数据集公之于众。

测试新电池设计的标准方法是充电和放电，直到它们失效。由于电池的寿命很长，这个过程可能需要数月甚至数年。
这项工作是在电池数据驱动设计中心进行的，这是一个集理论、实验和数据科学于一体的学术与工业合作机构。斯坦福大学的研究人员在材料科学与工程助理教授William Chueh的带领下进行了电池实验。由化学工程教授理查德·布拉茨(Richard Braatz)领导的麻省理工学院团队完成了机器学习的工作。该研究的共同主要作者克里斯汀·塞弗森(Kristen Severson)去年春天在麻省理工学院(MIT)完成了化学工程博士学位。

该项目的一个重点是找到一种能在10分钟内给电池充电的更好方法，这一功能可能会加速电动汽车的大规模普及。为了生成训练数据集，该团队对电池进行充放电，直到每一个电池的有效寿命结束，他们将其定义为容量损失20%。在优化快速充电的过程中，研究人员想要弄清楚是否有必要让电池在地面上充电。能否从早期循环的信息中找到电池问题的答案?

一般来说，锂离子电池的容量在一段时间内是稳定的。然后又急转直下。正如大多数21世纪的消费者所知道的那样，直线落点差异很大。在这个项目中，电池可以持续150到2300次。产生这种差异的部分原因是测试了不同的快速充电方法，但也与电池的制造差异有关。

这种新方法有许多潜在的应用。例如，它可以缩短验证新型电池的时间，这在材料快速进步的情况下尤其重要。有了这种分类技术，电动汽车电池的寿命被确定为很短——对汽车来说太短了——可以用来为路灯供电或为数据中心备份。回收者可以从使用过的电动汽车电池组中找到足够容量的电池，以供二次使用。

另一种可能是优化电池制造。“制造电池的最后一步叫做‘成型’，可能需要几天到几周的时间。

创新价值

这项技术不仅可以用于对制造的电池进行分类，还可以帮助新电池设计更快地进入市场。这一进展可能会加速电池的开发并改善制造。

创新关键点

结合综合实验数据和人工智能，揭示了在锂离子电池容量开始衰退之前准确预测其使用寿命的关键。在研究人员用数亿个电池充电和放电的数据点训练他们的机器学习模型后，算法根据电压下降和早期循环中的其他一些因素，预测每个电池还能持续多少次循环。

Use artificial intelligence to accurately predict the life of lithium-ion batteries

Scientists at Stanford University, the Massachusetts Institute of Technology and the Toyota Research Institute have found that a combination of synthetic experimental data and artificial intelligence reveals the key to accurately predicting the lifespan of lithium-ion batteries before their capacity begins to decline. After the researchers trained their machine learning models on hundreds of millions of battery charging and discharging data points, the algorithm predicted how many more cycles each battery would last, based on voltage dips and a number of other factors in earlier cycles.
These predictions were within 9% of the actual number of cycles the cells survived. In addition, the algorithm classifies batteries into long-life or short-life categories only based on the first five charge-discharge cycles. Here, the predictions were 95 percent accurate.
The machine-learning approach described in the journal Nature Energy could speed up research and development of new battery designs and reduce production time and costs, among other applications. Researchers have made this dataset, the largest of its kind, public.
The standard way to test new battery designs is to charge and discharge them until they fail. This process can take months or even years because of the battery's long life.
The work was carried out at the Battery Data-Driven Design Centre, an academic and industrial collaboration that combines theory, experiment and data science. The Stanford researchers, led by William Chueh, an assistant professor of materials science and engineering, conducted the battery experiment. The MIT team, led by chemical engineering professor Richard Braatz, did the machine learning work. Kristen Severson, the study's co-lead author, completed her doctorate in chemical engineering at MIT last spring.
One focus of the project is to find a better way to charge batteries in 10 minutes, a feature that could accelerate the mass adoption of electric vehicles. To generate the training dataset, the team charged and discharged the batteries until the end of each battery's useful life, which they defined as a 20 percent loss of capacity. In the process of optimizing fast charging, the researchers wanted to figure out if it was necessary to let the battery charge on the ground. Can the answers to the battery problems be found in the early loop information?
In general, the capacity of lithium-ion batteries is stable over a period of time. Then it took a turn for the worse. As most 21st century consumers know, the line drops vary widely. In this project, the battery lasts between 150 and 2,300 cycles. Part of the difference is due to testing different fast charging methods, but it also has to do with differences in battery manufacturing.
Attiya says the new method has many potential applications. For example, it could shorten the time it takes to verify new types of batteries, which is especially important when materials are advancing rapidly. With this sorting technology, EV batteries are determined to have a very short life - too short for a car - and can be used to power street lights or back up data centres. Recyclers can find enough capacity from a used EV battery pack for a second use.
Another possibility is to optimize battery manufacturing. "The final step to make a battery is called 'molding' and can take anywhere from a few days to a few weeks.