Computers Could Teach Us How to Build a Longer-Lasting Battery

How long will my battery last?

You’ve probably asked yourself this question innumerable times as you replaced the batteries in your cell phone or flashlight, or after purchasing an expensive new car battery. Having to guess at how long those batteries would continue powering your device or vehicle is not only inconvenient, but also expensive.

So far, battery researchers and manufacturers alike have had only one reliable way to test a battery’s lifespan—keep cycling it until the battery finally runs out of juice. ​Cycling a battery involves fully charging and then discharging it. Unfortunately, this method can take years and is expensive, electrochemist Susan ​Babinec says in a an Argonne National Laboratory press release about a new study she co-authored. This research takes advantage of machine learning, training computers to recognize patterns in data to make predictions about new data. In this case, computers were able to accurately predict how long different kinds of batteries will keep working.

Scientists at the U.S. Department of Energy’s Argonne National Laboratory in Lemont, Illinois gathered experimental data from 300 batteries representing six different chemistries, including the types and arrangements of atoms that make up the structure of the battery cathode. “Different cathode types can store more or less energy, and may degrade quicker or more slowly,” Noah Paulson, a computational scientist at Argonne and an author of the study, tells Popular Mechanics in an email. Another difference among the batteries was in chemical additives of the battery electrolyte.

The scientists let computers do the work to accurately determine just how long different batteries would continue to cycle. In this study, researchers studied lithium-ion batteries, which can be “charged and discharged thousands of times, depending on the way they are used,” Paulson says. This work reflects what would happen to most rechargeable batteries, such as lithium-ion, nickel-metal-hydride, or lead-acid, for example. “This is important as batteries for cars, planes, grid storage, electronics, and more must be rechargeable,” he explains. Alkaline batteries, on the other hand, like those you use in your TV remote, are typically not rechargeable.

The authors based their research on the fact that machine learning can predict lithium-ion battery lifespans in just a few weeks, from a maximum of 100 cycles. At the low end, the machine learning method the researchers developed took as little as one preliminary cycle to make a useful prediction, according to the study, published in the February 25 online edition of the Journal of Power Sources.

Batteries store chemical energy in the form of chemical compounds. For example, an alkaline battery, or cell, contains zinc, manganese dioxide, and potassium hydroxide. When attached at either end to a circuit such as a light bulb, the zinc inside reacts with the manganese dioxide and loses electrons. The electrons flow through a metal rod in the cell from the negative terminal to the bulb and make it light up. Then they continue flowing, and enter the positive terminal of the cell. In this type of non-rechargeable battery, once the zinc electrons are used up, the cell is dead.

Batteries powers our devices, tools, toys, and vehicles. No one knows how long a new battery will last, because many factors govern its potential lifespan—what we use them for, their internal chemistry, and their overall design.

“For every different kind of battery application, from cell phones to electric vehicles to grid storage, battery lifetime is of fundamental importance for every consumer,” Paulson says in the release. ​Testing a battery means running it through thousands of cycles as it converts its stored chemical energy into electricity. The machine learning method “creates a kind of computational test kitchen” that quickly reveals how a battery will perform, he explains.

To set up their experiment, the researchers fed the computers massive amounts of raw data, defining 397 distinct features of the batteries that they figured would be useful to the machine learning algorithms. The computers used a specific set of rules, or algorithms, to statistically analyze this raw “training data,” which was meant to familiarize the computers with the current and voltage versus time throughout the lives of the different batteries.

Based on the computers’ training, the machines learned to recognize patterns among various battery features and to build a model that could be used to make predictions about new data, such as average charging time for a battery. “Then we used an algorithm to find which subsets of the features would give the best predictions,” Paulson says.

By repeating this process, the computers’ predictions about new battery design became more accurate. “In this study, we only use this information from the first one to 100 charge-discharge cycles so that given a new battery, we can estimate its life without experimentally cycling it for months or years,” Paulson says.

The researchers based the machine learning algorithm on a lithium-based battery chemistry that’s well understood. They “trained” the computers on this algorithm to make predictions about the longevity of an unknown battery chemistry. ​“Essentially, the algorithm may help point us in the direction of new and improved chemistries that offer longer lifetimes,” Paulson says. The researchers believe that the machine could speed up the development of potential battery materials, because lab scientists would be able to test the material faster.

Different batteries degrade and fail in multiple ways, Paulson says. “The value of this study is that it gave us signals that are characteristic of how different batteries perform.” The most aggressive way to shorten a battery’s life is to charge it very quickly. As a result, lithium metal ends up coating the electrode particles. “Interactions between the electrolyte and electrode particles create a film that can protect the particles, but when it gets too thick it can become a barrier to [lithium] ions moving in and out of the particles,” Paulson says. Another way that battery life is cut short is through the cracking of electrode particles. Those are just a couple of examples.

“Say you have a new material, and you cycle it a few times. You could use our algorithm to predict its longevity, and then make decisions as to whether you want to continue to cycle it experimentally or not,” Paulson says.