Realizing the Promise of Early-Life Battery Failure Prediction Models with Active Learning

Since the first publication demonstrating early-life battery failure prediction by Severson et al [1], thousands of following research papers have demonstrated that usefully accurate battery lifetime predictions can be made using a variety of machine-learning techniques. The promise of these models is that they can accelerate battery development; the authors of [1] immediately followed their work with the results of Attia et al [2], demonstrating charge-protocol optimization using an early-life battery failure prediction model and active learning via a Bayesian optimization algorithm, resulting in a >10x decrease in the amount of testing hours required to reach a near optimal protocol versus testing batteries to failure and randomly searching for the charge-protocol (500 hours versus 7,700 [2]). However, as noted in that work, this effort required an existing data set to develop the early-life battery failure prediction model; the training data collected by Severson et al [1] totaled 81,487 hours of testing data from 124 cells, and prior work by Gasper et al [3] shows that, when randomly sampling, about 50% of that data (~60 cells tested to failure) is required to train a model with good predictive performance. However, battery research and development often relies on just a handful of cell tests to make decisions. So, to enable early-life battery failure prediction models to truly accelerate battery R&D, we need to demonstrate two critical developments: Demonstrate training of early-life battery failure prediction models from extremely limited initial data (< 10 cells), and Demonstrate active-learning exploration or optimization for an R&D goal by making uncertainty-aware decisions for both starting cells on new experimental protocols and stopping testing of existing cells when there is ‘enough’ data. If these two goals are satisfied, it should be feasible to both train usefully accurate early-life battery failure prediction models as well as accelerate the optimization or exploration of an R&D goal using dramatically less battery testing cost than has been traditionally required. This presentation will report on work related to both goals. For demonstration of early-life battery failure prediction models from limited data, we make apples-to-apples comparison of a variety of different model types across several open-source battery degradation data sets. We consider four types of models with inherently different architectures: ‘standard’ failure point models [1,2,3]; remaining useful life models; autoregressive trajectory models; and Bayesian trajectory models. Each model type is also augmented to enable them to learn both from intra-cell differences (changes in performance across a single cell’s lifetime) as well as inter-cell difference (changes in performance relative to other cells in a data set) [4]. Models are compared both with respect to failure prediction accuracy as well as confidence interval coverage accuracy, which quantifies the accuracy of the estimated uncertainty. These models are then utilized in an active learning framework, which has the goal of learning the relationship between an experimental variable (like charging protocol) to lifetime. Active learning starts from a small subset of cells, which are used to train any early-life battery failure prediction model as well as the experimental surrogate model. Then, iterative decisions are made to test new cells, recommending either the value of the experimental variable as well as the duration of test time estimated to result in a satisfactorily confident failure prediction (batch time recommendation), or recommending the value of the experimental variable and then choosing at each ‘cycle’ to stop or continue an aging test based on the early-life failure prediction model uncertainty (cycle-by-cycle recommendation). Performance of the active learning framework for both exploration and optimization tasks is demonstrated using open-source battery aging data sets. [1] Severson, K., et al. Nature Energy (2019) 4 383-391. [2] Attia, P., et al. Nature (2020) 578 397-402. [3] Gasper, P., et al, J. Electrochem. Soc. (2025) 172 043509. [4] Zhang, H., et al, arXiv:2310.05052v2 (2023). Figure 1