Scientists have developed a machine learning method that could significantly reduce the cost and energy needed to develop the new lithium-ion batteries that the modern world is increasingly relying on.
Predicting the lifespan of new battery designs and their engineering applications is a major bottleneck in the industry. Brute force testing a prototype, charging and discharging it repeatedly until it approaches an end-of-life threshold, can take months or even years, consumes enormous amounts of power, and is costly.
One study estimates that current and future lithium battery designs could require 130,000 GWh of energy from 2023 to 2040 if no changes are made to the development process. This is equivalent to about half of the annual electricity generated in California (278,338GWh).
A study published this week in the scientific journal Nature describes a new approach to machine learning in battery development, which the authors claim can save 98 percent of time and 95 percent of cost compared to traditional methods.
Chao Fu, an associate professor at the University of Connecticut, said in an accompanying article that this “shows great potential to address a major bottleneck in battery development.”
The process developed by University of Michigan postdoctoral researcher Jiawei Zhang and his team combines iterative elements to reduce the data needed to make accurate predictions.
The so-called Discovery Learning framework is considered highly accurate, building on a 2019 study that showed that machine learning models leveraging early life data from prototype battery tests could predict battery life with an average error of less than 15% on a test set.
Zhang et al. divided their previous method into three different components. The learner module selects new design prototypes that may provide useful data to improve prediction accuracy. After initial testing of these prototypes, the interpreter module uses models of physical properties to analyze this data along with historical full-life data for existing batteries. Finally, the Oracle module uses its output to predict the lifetime of the newly tested prototype. Importantly, that information is fed back to the learner module to select the next set of prototypes to physically test.
“The key novelty of the Discovery Learning model is that it can update itself using Oracle-predicted lifetimes rather than experimentally measured lifetimes, thereby avoiding the need for time-consuming full-life battery testing,” Hu said.
However, he notes that it remains unclear how well the Discovery Learning framework will perform if new battery designs deviate significantly from those of batteries available to provide training data.
“More broadly, before this framework is adopted for general use, further validation is required to see how well it withstands batteries used in real-world conditions, for example at different temperatures and under different electrical loads,” Hu said.
Nevertheless, with the global value of batteries for EVs, laptops, and various other applications currently reaching $120 billion and expected to rise to nearly $500 billion by 2030, even small savings in development costs can make a difference. ®
