New ML-based Approach for Faster Battery Testing

Energy storage is an inseparable part of many applications, including the electrical grid, electric vehicles, portable electronics, and many more. Recently, there has been much research on energy storage due to its contribution to the electrified future. However, an essential task of battery research and development is evaluating and optimizing battery performance.

Optimization and evaluation are costly processes as the battery cycling is time-consuming. Waiting to collect the battery data after thousands of charge and discharge cycles for optimization is not feasible. Moreover, if the optimization parameters are more, many trials are required to explore different parameter configurations.

Researchers at the University of Michigan introduce a generic framework leveraging machine learning (ML) algorithms to optimize battery parameters and improve battery performance efficiently.

 

Current Methods to Evaluate Battery Performance

Battery degradation is a continuous process resulting from various chemical and mechanical changes to the electrodes. Some prominent reasons for the degradation of lithium-ion batteries are the growth of resistive layers called solid electrolyte interphase (SEI) on the electrode surface, thermal decomposition of electrolyte, and mechanical cracking of the SEI layer.

The SEI layer is a passivation coating formed by electrolyte degradation products and is responsible for improved performance as it stabilizes the anode-electrolyte interface. However, it also consumes lithium ions, reducing the overall charge and discharge efficiency of the electrode material. Therefore, lithium-ion batteries are prone to capacity fading after hundreds of cycles.

There are many methods to evaluate battery performance for optimization. One of the easiest and quickest ways is to use simple equivalent circuit models and regression models to optimize the objective by algorithms such as linear programming. However, the accuracy of this method is limited due to the simplified models. For example, many models approximate SEI growth to be proportional to the square root of time, which is not unjustifiable in many cases. More accurate methods rely on physical-based pseudo-two-dimensional models. However, they are time-consuming and unreliable for new battery chemistries. Therefore, more reliable methods are necessary when dealing with upcoming battery chemistries.

 

New Approach to Battery Testing

To significantly reduce the testing time and cost, researchers at the University of Michigan take advantage of the latest machine learning algorithms to create a system that decides when to quit testing and how to get better at it. The new optimization system reduces the time for simulation and physical testing of new batteries by about 75%.

The algorithm stops cycling tests when the battery doesn't show promising results at the start to save the resources. The framework uses mathematical techniques known as Asynchronous Successive Halving Algorithm and Hyperband.

The approach consists of a pruner and a sampler. The pruner, using mathematical techniques, halts the unpromising cycling batteries to reduce costs. Using the Tree of Parzen Estimator, the sampler predicts the new set of promising parameters based on the tests.

In addition, the new system generates multiple battery configurations to be tested parallelly. During tests, if any configuration completes testing or shows unpromising results, new configurations are calculated without waiting for the results of other tests.

 

Illustration of the new battery testing technique. Image used courtesy of the University of Michigan

"This framework can be tuned to be more efficient when a performance prediction model is incorporated," said Changyu Deng, U-M mechanical engineering doctoral student and the paper's first author. "We expect this work to inspire improved methods that lead us to optimal batteries to make better EVs and other life-improving devices.

 

Feature image used courtesy of Pixabay