Data-driven prediction of battery cycle life before capacity degradation

Abstract

Accurately predicting lifetime of complex systems like lithium-ion batteries is crucial for accelerating technology development. However, diverse aging 1 mechanisms, significant device variability, and varied operating conditions have remained major challenges. To study this problem, we generated a dataset consisting of 124 commercial lithium-iron-phosphate/graphite cells cycled under fast charging conditions. The cells exhibited widely varied cycle lives spanning from 150 to 2,300 cycles, with end-of-life defined as 20% degradation from nominal capacity. Using discharge voltage curves from early cycles yet to exhibit capacity degradation, we apply machine learning tools to predict cycle life with less than 15% error on average, which is improved to ~8% error by incorporating additional data. Our work represents a significant improvement over previous predictions that generally required data corresponding to >5% capacity degradation, without needing specialized diagnostics. Additionally, it highlights the promise of combining data generation with data-driven modeling to predict the behavior of complex and variable systems. Main Lithium-ion batteries are deployed in a wide range of applications due to their low and falling costs, high energy densities, and long cycle lives. 1-3 However, as is the case with many chemical, mechanical, and electronics systems, long battery cycle life implies delayed feedback of performance during development and manufacture, often many months to years. Accurately predicting cycle life using early-cycle data would accelerate this feedback loop as well as enable estimation of battery life expectancy for use in consumer electronics, electric vehicles, and second-life applications. 4-6

Predicting battery life with early cyclic data by machine learning

“…In this work, the database consists of 138 sets of data recorded on the commercial LFP/graphite A123 APR18650M1A cells. The nominal capacity of these cells is 1.1 Ah, and their nominal voltage is 3.3 V. These cells were tested in the Arbin LBT potentiostat at 30°C …”

“…Accordingly, the battery was divided into two groups: “High lifetime” and “Low lifetime”. The criterion for determining battery failure is that the cycle capacity is less than 80% of the initial capacity . “High lifetime” batteries can cycle more than 550 cycles before battery failure.…”

“…Recently, with the development of battery‐related technologies, more and more data of these energy storage devices have been produced. For example, a database using commercial lithium iron phosphate (LFP)‐graphite batteries were established . Based on these accumulated data, data processing methods, such as machine learning, have been applied gradually to analyze the properties of the battery.…”

“…3,9 However, these methods have encountered some problems, the most important of which is the nonlinear degradation of battery with cycling and extensive variability. 10,11 In complex systems, the applicability of semi-empirical formula or physical/ chemical analysis is limited to a certain range. Instead, the complex battery system requires more advanced statistical or mathematical methods for analysis.…”

“…For example, a database using commercial lithium iron phosphate (LFP)-graphite batteries were established. 11,12 Based on these accumulated data, data processing methods, such as machine learning, have been applied gradually to analyze the properties of the battery. Wang et al applied support vector machines to build a battery life prediction framework.…”