A new artificial intelligence model accurately predicts the lifetime of lithium-ion batteries.
By using the model, a data set of approximately 96,700 cycles was prepared that evaluates lithium-ion batteries cycled under controlled conditions.
Under development for some time,
lithium-ion batteries have come under scrutiny due to performance deficiencies. One remaining issue is the formation of spiky structures known as dendrites that originate at the battery’s anode. If left unchecked, dendrites will grow until they reach the battery’s cathode, leading to a short circuit that can cause the battery to overheat and potentially catch fire.
In a previous TLT article, researchers working with a lithium-metal anode in a lithium-sulfur battery suppressed dendrite formation through the deposition of multi-walled carbon nanotubes on the anode (1
). Lithiation of the carbon nanotubes appears to prevent dendrite formation by enhancing the solid electrolyte interphase formed between the anode and the electrolyte.
The uncertainty surrounding the performance of lithium-ion batteries leads to the question about whether any research has been done to accurately predict battery life. Peter Attia, a doctoral candidate working with William Chueh, assistant professor of material science and engineering and Center Fellow at the Precourt Institute for Energy at Stanford University in Palo Alto, Calif., says, “Two main approaches have been taken in trying to predict the life of a battery. Physics-based models have been proposed to account for degradation modes such as lithium plating, loss of active material and impedance increase. The modeling has evaluated batteries that were in an isolated condition and often undergoing slow cycling. This approach was not an accurate predictor of battery life and represented evaluating a battery placed on a shelf.”
The second approach was data driven and represented a semi-empirical strategy. Attia says, “Predicting battery life using this method had two limitations. The first one is that one of the most common publicly available data set is developed on four batteries. Second, extracting all of the data from battery recyclers can be tedious and quite challenging.”
With the improvement in computational power and data-generation techniques, researchers now have more resources to develop a better strategy for predicting battery life. Such an approach using artificial intelligence has been developed.
Attia, Chueh and their collaborators developed an artificial intelligence model that accurately predicts the lifetime of lithium-ion batteries (see Figure 1
). Attia says, “Our approach was to initially generate a data set containing the results from fast-charging 124 commercial batteries that underwent varying charge/discharge cycles. We set up a battery cycler that can run 48 tests in parallel to produce the results. The cells have cycle lives ranging from 150 to 2,300 cycles.”
Figure 1. A new artificial intelligence approach accurately predicts the cycle life of a lithium-ion battery using parameters that evaluate performance. (Figure courtesy of Stanford University.)
The batteries were discharged under 4 C conditions, which amounted to taking around 15 minutes to discharge a battery. In evaluating battery performance, the researchers determined that failure occurred when a battery reached 80% of its normal capacity.
In generating this data, the researchers compiled a data set with approximately 96,700 cycles that they claim is the largest publicly available lithium-ion battery set available on commercial lithium-ion batteries cycled under controlled conditions. The batteries contained lithium iron phosphate cathodes and graphite anodes.
Attia says, “The biggest limitation on battery life was on the graphite side. The cathodes did not degrade during our testing. This meant that we could focus our modeling on predicting how quickly the anode degraded.”
The large range of battery cycle lives provided the researchers with the opportunity to design their machine-learning model to accurately predict failure. Attia says, “We used features that explain how lithium-ion batteries perform in developing models. Our features included building three different models. The first model evaluated changes in the voltage of batteries as a function of the number of cycles. In the second model, we examined changes in the battery’s voltage and capacity as a function of the number of cycles. For the third model, we evaluated changes in battery capacity, voltage, temperature and internal resistance as a function of the number of cycles.”
Attia indicated that the most accurate model in predicting battery life was the third model but that the most important parameter is voltage. He says, “Using the machine learning tools we developed enabled us to accurately predict battery life with an error of only 9.1% using data produced during the battery’s first 100 cycles. The important feature is that we were able to evaluate lithium-ion batteries under realistic operating conditions and determine very accurately when a specific battery will fail well before the battery shows any indication of failing.”
Attia believes that this machine-learning approach can be used in other applications. He says, “We feel this technique can be applied to other types of batteries, including those used in automobiles. Another application is to expedite the production of batteries. The last step in battery production is formation cycling and is needed to form a stable solid electrolyte interphase on the anode. This step is expensive and can sometimes take weeks to complete.”
A third potential application is to predict the remaining life of batteries that have already been used once. Attia says, “Determining an optimum second-life application for a specific battery will be a good way to maximize value through recycling.”
Further information can be found in a recent article (2
) or by contacting Chueh at firstname.lastname@example.org
Canter, N. (2019), “Minimization of Dendrite Formation in Lithium-Sulfur Batteries,” TLT, 75
(2), pp. 10-11.
Severson, K., Attia, P., Jin, N., Perkins N., Jiang, B., Yang, Z., Chen, M., Aykol, M., Herring, P., Fraggedakis, D., Bazant, M., Harris, S., Chueh, W. and Braatz, R. (2019), “Data-Driven Prediction of Battery Cycle Life Before Capacity Degradation,” Nature Energy
(5), pp. 383-391.