AI Is the New Secret Weapon in the Quest for Better Batteries

Illustration for article titled AI Is the New Secret Weapon in the Quest for Better Batteries
Photo: Sam Rutherford (Gizmodo)

Compared to all the electronics that power the tiny computer in your pocket; battery technology is downright disappointing. Not only does your smartphone need charging every day, but in a few years, its battery will be barely able to hold a charge at all. So how long will your device last? Researchers at Standford University and MIT have created an AI that can predict a battery’s potential lifespan after just a handful of charges.

Compared to the rechargeable battery technology that existed even just 20 years ago, the lithium-ion battery inside your phone, tablet, and almost any mobile device is a marked improvement. They’re now relatively cheap to manufacture, and they tend to provide solid performance right up until the end of their life cycle. But knowing when that end is coming has typically been very hard to predict, which has also made the research and development of new battery technology very time-consuming.

Will a different combination of chemicals and materials result in a lithium-ion battery variant that lasts much longer than its predecessors? Does this new approach to fast-charging have longterm consequences for the life of the battery? The only way to tell is to repeatedly recharge and discharge a sample until it reaches the end of its life cycle, which is defined as having less than 20 percent of its original power capacity. That’s a time-consuming process, and it’s partly why it feels like innovations in battery tech haven’t kept pace with electronics tech.


To potentially help accelerate battery R&D, the researchers at Stanford University and MIT worked with the Toyota Research Institute and used machine learning to develop an algorithm that can very accurately predict a battery’s performance. Trained on hundreds of millions of measurements gathered while batteries were being charged and drained—including power capacity, charging times, and even the temperatures of the battery cells—the algorithm can predict how many cycles a battery can be effectively charged and discharged, with actual charging results being within nine percent of the prediction. The algorithm doesn’t completely replace actually testing samples until they die, but it could help engineers quickly ascertain if changes they’re testing have the potential for improvement.

Even with stringent manufacturing tolerances, the lifespan of one battery can be different than the one that rolled off the assembly line before it. Using the same machine learning algorithm, the researchers found that 95 percent of the time they were able to accurately predict if a battery would have a longer or shorter lifespan after gathering sample data from just five charging cycles. That approach doesn’t predict exactly how long a battery will last, but it could allow manufacturers to more efficiently sort batteries as they roll off the assembly line, so the ones with a much longer predicted lifespan could be reserved for smartphones or electric cars that have much higher power demands.

[Nature Energy via EurekAlert!]

Share This Story

Get our `newsletter`


Dense non aqueous phase liquid

Chemists and chemical engineers do this. Then again, the article is about speeding up repetitive bench chemistry for a better battery. This sort of falls under the general methods of inductive analysis (it’s better to have a machine for that). 

Here’s a breezy paper on the subject:

Learning to Predict Chemical Reactions

A snippet for the gist (from abstract):

Being able to predict the course of arbitrary chemical reactions is essential to the theory and applications of organic chemistry. Approaches to the reaction prediction problems can be organized around three poles corresponding to: (1) physical laws; (2) rule-based expert systems; and (3) inductive machine learning. Previous approaches at these poles, respectively, are not high throughput, are not generalizable or scalable, and lack sufficient data and structure to be implemented. We propose a new approach to reaction prediction utilizing elements from each pole. Using a physically inspired conceptualization, we describe single mechanistic reactions as interactions between coarse approximations of molecular orbitals (MOs) and use topological and physicochemical attributes as descriptors.

End of chemistry crap.