Artificial Intelligence

New AI Method Speeds Battery Designs for Electric Cars

Stanford-led AI research uses machine learning to reduce battery testing times.

Posted February 22, 2020 | Reviewed by Davia Sills

Source: paulbr75/Pixabay

The life of lithium-ion batteries is both a bane and blessing of modern-day living, as many people depend on it each day to power devices such as mobile phones, laptops, and hybrid and electric cars. By 2025, the lithium-ion battery market is projected to reach $93.1 billion, growing at a CAGR of 17 percent, according to Grand View Research. As the demand for electric cars and rechargeable electronic devices increases, so will the need for the development of higher-performance lithium-ion batteries.

In a recent Stanford News report, a study led by Stanford University professors Stefano Ermon and William Chueh, along with their colleagues at the Massachusetts Institute of Technology (MIT) and the Toyota Research Institute, created a battery testing method using artificial intelligence (AI) machine learning that reduces battery testing times by 98 percent.

Published in Nature on February 19, 2020, the research study showed how the combination of an early-prediction model with a Bayesian optimization algorithm can decrease the battery testing process from nearly two years to 16 days.

Machine learning algorithms depend on the effective calibration of learning parameters and the model hyperparameters that guide the overall training process. However, the tuning is more of a form of art as it may depend on heuristics, domain expert experience, or time-consuming search.

Bayesian optimization, based on the Bayes Theorem, is well suited in machine learning to automatically tune the hyperparameters for time and cost-efficiency. It finds the global maxima of a black-box function that is most conducive to a favorable outcome, using the minimum number of steps.

In this research study, first, the team created an early prediction model based solely on a few charging cycles that could predict how batteries react to various charging protocols and could identify in real-time which ones were desirable.

The team developed a Bayesian optimization algorithm that decreased the number of experiments needed to explore the parameter space of charging protocols with the purpose of identifying the optimal way to charge an electric vehicle battery in 10 minutes that maximizes the overall life of the battery. The early prediction model is able to produce predictions based on data from a few cycles, thereby reducing the time per experiment.

The method developed accelerates the testing process for ultra-fast battery charging. It automatically incorporates feedback from past experiments in order to provide future decision-making guidance. According to the researchers, not only can their method be adapted to perform other processes in the battery design and development process, but also it can be used more widely for any scientific domain that requires multi-dimensional design spaces and time-intensive experiments.