Optimizing Battery Lifetime under Fast Charging Conditions through Machine Learning and Experimental Design

By Kent Griffith

March 3, 2020 | High-throughput and accelerated aging tests have been used to try to tackle the numerous variables and long experimental times in battery R&D. Accelerated aging typically involves cycling the batteries at a higher temperature and/or faster rate than they would normally face in practical conditions. Rather than isolated experiments, it is possible to gain insights into battery operation and failure through closed-loop optimal experimental design approaches that feed back the results from a test into the next round of testing.

In an article published last week in Nature (DOI: 10.1038/s41586-020-1994-5), a team of materials scientists, chemical engineers, and computer scientists from Stanford, MIT, Toyota Research Institute, Lawrence Berkeley National Laboratory, and SLAC National Accelerator Laboratory used this strategy to develop a machine-learning-based method for optimizing fast-charging protocols with fewer experiments and an accelerated timeline.

Starting with commercial lithium iron phosphate (LFP)//graphite 1.1 Ah cylindrical cells from A123 Systems, the authors sought to identify a charging protocol that would maximize the cycle life. Fast charging can lead to degradation of the delicate surface interphase layer (SEI) on the graphite anode, delamination of the graphite layers, or the deposition of dendrites—metallic lithium fibers—on the surface of the graphite. At the extreme, dendrites can cause short-circuits and lead to battery fires; under milder conditions, they can still lead to degradation of the battery cell through the formation of “dead lithium” that is lost from the electrode reservoirs. Fast charging also creates inhomogeneities within a battery that lead to degradation when a portion of an electrode within a cell is overcharged. The same mechanism can happen to a subset of cells within a battery pack.

It was known that simply applying one large constant current is not the most effective way to perform fast charging; however, the best fast charging protocol has not been established. In the present study, the team examined 224 different charging protocols. Each protocol consisted of six steps; the first five were at different constant currents and the sixth was at constant voltage. For reference, standard commercial chargers often use a single constant current up to 70–80% state-of-charge, followed by a constant voltage step to full charge, which prevents overcharging. In this new report, the first three steps were independent within a C-rate range of about 4–8C (1C = 1.1 A for the 1.1 Ah cells); step four was set to ensure a charging time of 10 minutes to reach 80% state-of-charge; step five was a further 1C constant current to 90% state-of-charge; and the final step was a constant voltage hold at 3.6 V.

After the closed-loop optimization, the 224 fast charging protocols led to a predicted cycle life of nearly 1200 for the best protocol but only about 600 for the worst-performing protocol. This highlights the significance of the charging protocol, potentially doubling cell lifetime. In this work, in accordance with standard practice, cycle life is defined as the cycle number when the battery retains only 80% of its original capacity. The results of this study showed that the top two fast-charging protocols used a 5.0±0.2C step for each of the three initial, independent charging steps, which is a potentially surprising outcome considering that the literature has been focused on protocols that decrease the charge rate as a function of state-of-charge (e.g. 8C–6C–4C). A key to the efficiency of this approach was the ability to predict cycle life from just the first 100 cycles, which was reported last year by the team. As validation of the cycle life prediction and closed-loop optimization, a subset of cells was cycled to end-of-life using the fast charging protocols identified here by machine learning and from the literature. There was a strong correlation in the early-estimated vs. actual cycle life using all protocols, and the predicted protocols were the top performing of all those measured.

It is worth noting that the authors chose LFP//graphite cells because LFP is believed to experience minimal degradation and thus the role of graphite would be relatively isolated, which is important since graphite is the dominant anode material. However, the limitations of graphite as a truly high-rate anode are well-established. Even under optimal conditions and a charge time of 10 mins to 80% state-of-charge and 16 mins to 90% state-of-charge, the LFP//graphite cells here reached end-of-life at a modest ~1000 cycles. Application of this methodology to fast-charging protocols for a cell with e.g. commercial lithium titanate (LTO) or emerging high-rate anodes would be an interesting future direction.

While this work focused solely on optimizing the charging protocol of a single cell type with a defined rate target, the authors note that future studies could extend this approach to include the design of battery materials and manufacturing processes.

Professor David Howey from the University of Oxford, who was not involved in this study, notes that the power of this approach is in the combination of lifetime prediction and design of experiments. He comments that this really showcases the utility of data-driven approaches while he also emphasizes that machine learning is limited by its training dataset and there can be important factors that might be better captured in a physics-based model. Both Howey and the authors of the study point out that the optimal charging protocol discovered here is not necessarily transferable to other cell chemistries, temperatures, voltage windows, etc. and that the emphasis is in the generality not of the results but of the methodology.

For those interested in applying this methodology to their own projects, the authors have put the code, data, and data processing algorithms online at https://github.com/chueh-ermon/battery-fast-charging-optimization, https://github.com/chueh-ermon/BMS-autoanalysis, and https://github.com/chueh-ermon/automate-Arbin-schedule-file-creation. Several of the authors have filed a patent on the basis of this work: US Patent Application No. 16/161,790 (16 October 2018).