A Framework for Optimal Safety Li-ion Batteries Design using Physics-Based Models and Machine Learning Approaches

Abstract

Numerical physics-based models for Li-ion batteries under abuse conditions are useful in understanding failure mechanisms and deciding safety designs. Since battery design is generally required to decrease the failure risks while increasing the performance, multi-objective optimization methods are useful. Nevertheless, these usually require huge computational costs because these models targeting abuse battery conditions generally have many input physical parameters and computational costs for calculating one result are high. Therefore, we develop a framework for performing multi-objective optimization at a reasonable computational cost using machine learning methods. With this framework, an inverse analysis of optimal Li-ion battery design conditions, including safety conditions, is performed. Nail penetration simulations on different input conditions are performed so as to build a database for battery design conditions/test conditions (descriptors) and safety/performance (predictors). As a result of analyzing the relationship between descriptors and predictors, a high correlation between fire spread and negative electrode active material diameter is confirmed. Furthermore, a regression model to predict the database is created with a Gaussian process model. Using the model and a genetic algorithm, optimal design conditions are searched, and the design conditions that offer higher safety and better performance are identified under the assumed conditions.