Battery energy storage (BES) is undergoing prolific growth into new areas and within existing areas such as vehicles and stationary scenarios.¹ Over the next decade, there will be a significant wave of vehicle electrification to replace conventional vehicle that are powered by fossil fuels.² With the even wider deployment of BES, an improved battery management with real time diagnostic capabilities and life prediction becomes necessary, allowing end-users to adjust usage strategies.³ To industry and in research there is tremendous economic and technical benefit to shortening the test period of batteries through robust predictive methods. Accurate long-term forecasting of battery life allows for more proactive planning of battery management (e.g., cell replacements) and preemptive actions of modified operating conditions to achieve safe conditions and prolong battery life. The ever-evolving landscape of battery materials (higher voltage, higher rate capable, greater energy storage) and applications (fast charge, low temperature, high temperature) ensure that there will be an abiding need for early capture of aging mechanisms.

The primary target for this study is accelerated identification of battery aging under extreme fast charge (XFC) conditions, wherein there can be a combination of foremost mechanisms under three broad “aging modes”: loss of lithium inventory (LLI), and loss of active materials in the positive and negative electrodes (LAM_PE and LAM_NE). Sigmoidal rate expressions (SREs) are employed as an accelerated diagnostic and predictive method to evaluate the aging mode. SREs are robust forms that contain three variables that capture the thermodynamic and kinetic “thumbprint” of the mechanism progression within the context of a batch reactor scenario.⁴ We demonstrate three different methods by which SRE parameters can be early assessed; (1) extrapolative techniques using specialized functions to determine SRE parameter convergence, (2) a technique based on deep learning and Monte Carlo framework, and (3) a technique based on Physics-Guided machine learning. Our analyses are based on data obtained through XFC testing of prototypical Li-ion cells built with NMC532 cathodes and graphitic anodes. Overall results from the three methods indicate that for the LLI-dominant cases we predict capacity loss at the end of test (typically after 450-600 cycles, 9-12 weeks) by using only Reference Performance Test data within the first 2-3 weeks. In many cases, the aging mode predictive error is within 5-10% relative error and 1-2% absolute error.

References

1. U.S. Department of Energy, Energy Storage Grand Challenge Market Report (2020).
2. D. Gielen et al. Energy Strategy Reviews, 24, 38-50 (2019).
3. C.R. Birkl et al., J. Power Sources, 341, 373–386 (2017).
4. K. Gering, Electrochimica Acta, 228(20), 636-651 (2017).