A generic battery-cycling optimization framework with learned sampling and early stopping strategies.

Area of Science:

• Battery Technology
• Machine Learning Applications
• Computational Science

Background:

• Battery optimization is hindered by high costs and time investment for experimental or simulation-based evaluations.
• Optimizing battery cycling performance is particularly time-intensive due to the lengthy nature of battery cycling processes.

Purpose of the Study:

• To develop an efficient machine learning framework for optimizing battery parameters.
• To significantly reduce the total battery cycling time and associated costs.
• To accelerate the exploration of battery configurations for both simulations and experiments.

Main Methods:

• The framework integrates a pruner utilizing the Asynchronous Successive Halving Algorithm and Hyperband to eliminate unpromising battery cycling.
• A sampler employing Tree of Parzen Estimators predicts the most promising configurations based on historical data.
• The system supports categorical, discrete, and continuous parameters and operates asynchronously in parallel.

Main Results:

• Demonstrated effective performance on a parameter-fitting problem for calendar aging.
• The framework successfully reduced the time and cost associated with battery optimization.
• Enabled efficient exploration of multiple simultaneous cycling cells.

Conclusions:

• The developed machine learning framework offers an efficient solution for battery optimization challenges.
• It significantly reduces the time and cost of evaluating battery configurations.
• The framework is applicable to both experimental and simulation-based battery research, fostering advancements in the field.