A battery’s design, combined with its cycling protocol, together determine its usable lifetime. Quantitatively predicting this interplay is an objective of enormous value in both research and deployment. Unfortunately, the extensive resources and time required for the analysis today result in a lack of detailed information on degradation curves; this challenge is currently handled by degradation estimates, which introduce deployment risks and hamper design of creative solutions to complex situations. Various intersecting degradation behaviors of both materials and subcomponents are responsible for the enormous complexity. 

During use, interior material degradation and structural changes modify the battery state of health (SOH); the present definition of SOH as % capacity retention does not predict future behavior and therefore inadequately represents the actual design state (“health”). For example, three cells with the same cycling protocol may retain the same % capacity at some standard number of cycles, but have drastically different future behaviors and different true SOH at that point. Existing predictive approaches have fallen short. Mechanistic chemistry- and physics-based models can describe the complex, nonlinear and mutually intersecting failure modes, but the difficulty of obtaining microstructure details and physical constants to inform the models hinders their predictive capabilities. This challenge motivates the use of advanced analytics (AI/ML) to enhance or replace the mechanistic models. Argonne’s approach, based on advanced state of health (A-SOH) descriptors, will be more quantitatively predictive and require only limited early cycle performance data to predict future battery performance.

To this end, an Argonne team of data scientists and electrochemists are developing a comprehensive AI-driven approach to diagnose the health of Li-ion batteries with limited usage data and predict battery life and degradation trends. A critical component of this approach is obtaining useful battery degradation data with which to train the algorithms. The team has adopted a parallel strategy: collecting historical testing data from Argonne and beyond, totaling thousands of cells, and generating synthetic cycling data through a range of physics-based approaches. To make best use of this data, the team explored physics-informed features tailored to both traditional and neural-network-based ML predictors. Using features extracted from the first 10-100 cycles of battery usage, deep learning predictors (e.g., recurrent neural networks) can accurately predict the degradation behavior of a previously unseen cell with a variety of chemistries and usage scenarios. Further, unsupervised ML algorithms can help to evaluate the similarity of a given cell to others on the basis of the cathode/anode/electrolyte composition, cycling protocol, environmental temperature and more. This game-changing capability could help accelerate the development of new battery chemistries and help commercial entities make the best use of their cells.