Machine Learning Assisted Discovery of Novel Sodium-Ion Battery Materials
This story is contributed by Sedzro Tamakloe
- Sodium-ion batteries (SIBs) will be a key technology to meet battery production needs due to the abundance of sodium, the low cost of the materials, and the elimination of cobalt from the cathode.
- Although SIBs can not match nickel-based lithium ion batteries (LIBs) in terms of energy density, CATL has produced a SIB that is comparable in energy density to lithium iron phosphate (LFP).
- Researchers have started to use machine learning (ML) to rapidly identify new SIB materials with comparable properties to LIB.
Introduction
Current state of SIBs
Sodium-ion batteries (SIBs) have the potential to become an attractive alternative to lithium-ion batteries (LIBs). Due to the abundance of sodium, the sixth most abundant material in the Earth’s crust, SIBs are much cheaper to make than LIBs. In addition, SIBs do not typically contain cobalt in the cathode. The lack of cobalt combined with the abundance and low cost of sodium make SIBs a promising option for energy storage.
Commercially, SIBs are well suited for energy storage applications such as balancing out grid renewables due to their wide temperature of operation and excellent cycle life. However, in applications that require high energy density, LIBs are generally superior. Contemporary Amperex Technology (CATL) has recently released an SIB that is slightly more energy dense than the LFP chemistry (160 Wh/kg). Although CATL’s SIB performs on par with LFP in terms of energy density, it cannot match the nickel cobalt aluminum chemistry (285 Wh/kg) used in current Tesla models.
Future Outlook and Research Direction
Despite the disadvantage in energy density, the development of SIBs to meet battery demand could be crucial in the next ten years. According to Rystad Energy, annual battery production is expected to exceed 7 TWh by 2030, equivalent to production from 200 Tesla gigafactories [1]. One of the main factors contributing to this ramp in production is the transition from internal combustion engine (ICE) vehicles to electric vehicles (EVs) from companies like Audi and GM, which will stop producing ICE vehicles in 2026 and 2035 respectively. In 2020, only 0.45 TWh of battery production was produced worldwide [2]. To meet the forecasted production by the end of the decade, LIB supply will need to increase substantially.
Machine Learning to Identify Novel Sodium-ion Materials
As described in Dr. Milan Sadan’s article, SIBs have the potential to be the future of energy storage. However, experimental approaches to synthesize breakthrough materials for SIBs have progressed slowly in terms of identifying chemistries that are comparable to LIBs for portable and high-energy density applications. The experimental approach generally consists of a structure-property relationship where the researcher tests a structure to achieve a desired property. If this test is successful, then the experiment is finished; if not, then the experimenters will modify the structure and iterate on the process until they find a material that provides the desired properties. This has been the framework for most new materials development, including that of SIBs.
To accelerate the discovery of novel SIB materials, researchers have adapted machine learning (ML) approaches to intelligently and rapidly predict SIB materials that might have superior properties. The ML approach uses all of the data currently available in the literature to predict structures that might yield the desired property. This method enables researchers to evaluate more structures rapidly and explore novel compositions they might not have otherwise considered.
ML for LIB Discovery
Numerous studies have used ML to discover novel LIB materials. For example, the Evan Reed group at Stanford came up with a novel LIB ionic conductor from a set of 12000 lithium-containing solid materials from the Materials Project database [3]. Their training set consisted of 40 well-known lithium solid electrolytes with 20 computed features for each material based on structural information and chemistry, such as electronegativity and bond number. Density functional theory (DFT) confirmed that an electrolyte candidate based on the ML model’s prediction from the lithium thioborote family of solid conductors showed superionic conduction and improved electrochemical stability over lithium thiophosphate materials. The group at Waterloo University led by Linda Nazar recently synthesized this material with halogen dopants and indeed found a superionic conductivity of 1.4 *10–3 S/cm for Li2B10S18 X1.5 (X = Cl, Br, and I) [4].
Another study from researchers at Purdue University showed that the best performing lithium lanthanum zirconium oxide (LLZO) material could have been found utilizing only 30% of the experimental studies used to date [5]. Verduzco et al. developed an active learning (definition) approach where they compare four acquisition functions: maximum likelihood of improvement, maximum expected improvement, upper confidence bound, and maximum uncertainty vs random guessing to see how many tries it would take to identify the top ionic conductor in a training set of LLZO ceramic electrolytes. All the acquisition functions were shown to identify the top ionic conductor in about 30% of the experiments required by random guessing proving active learning’s capability as a design tool to discover novel battery materials.
ML for SIB discovery
ML has also been applied to the discovery of novel materials for SIBs. Venkat Viswanathan, a professor in Carnegie Mellon University’s mechanical engineering department, developed an autonomous approach that combines robotics and ML to perform hundreds of sequential experiments to optimize battery electrolytes [6]. The training set consists of 251 existing aqueous electrolytes with features such as conductivity and pH. Using an algorithm that runs without human intervention, a novel, non-intuitive aqueous sodium-ion electrolyte was discovered with an improved chemical stability window compared to the feeder solution benchmark. This work was the first automated design of battery electrolytes coupling ML with robotics and offers a path forward for the discovery of more materials for SIBs.
Moses et al. developed a deep neural network regression model trained on data from the Materials Project to predict materials with a high average voltage and low volume expansion during charge and discharge [7]. After training, the model was applied to a test set of new SIB electrodes with good performance in energy density and small volume change upon charge and discharge, as predicted by ML. The model’s predictions were then compared to DFT to assess the model’s viability in predicting novel materials for battery applications. The neural network had a relatively lower mean absolute error than the DFT calculated results for percent volume expansion and average voltage, further demonstrating the role of ML in exploring novel materials.
Conclusion
With further advances in research and industry, SIBs have the potential to play a major role in energy storage, especially due to the cheap and abundant nature of sodium. Researchers are beginning to utilize ML techniques to identify novel SIB materials at an expedited rate. The implementation of ML for battery electrodes has led to the identification of the top candidates in the design space at a fraction of the effort needed to do so experimentally. Further optimization of early stage R&D through machine learning will lead to the development of better materials for a wide range of energy storage applications.
Sedzro Tamakloe is a PhD researcher at The Ohio State University in materials science and engineering. His work focuses on utilizing machine learning and statistical approaches to discover and improve electrodes for sodium-ion batteries. He enjoys fellowship with friends, tennis and serving others.
Interested in publishing in BatteryBits? Apply at this link to become a contributor.
Join our discussion with other battery professionals at the community slack space Battery Street!
References
[3] Sendek, A. D., Yang, Q., Cubuk, E. D., Duerloo, K. A. N., Cui, Y., & Reed, E. J. (2017). Holistic computational structure screening of more than 12 000 candidates for solid lithium-ion conductor materials. Energy and Environmental Science, 10(1), 306–320.
[4] Kaup, Kavish, Assoud, Abdeljalil, Liu, Jue, and Nazar, Linda F. Fast Li-Ion Conductivity in Superadamantanoid Lithium Thioborate Halides. United States: N. p., 2020.
[5] Verduzco, J. C., Marinero, E. E., & Strachan, A. (2021). An Active Learning Approach for the Design of Doped LLZO Ceramic Garnets for Battery Applications. Integrating Materials and Manufacturing Innovation, 10(2), 299–310.
[6] Dave, A., Mitchell, J., Kandasamy, K., Wang, H., Burke, S., Paria, B., Póczos, B., Whitacre, J., & Viswanathan, V. (2020). Autonomous Discovery of Battery Electrolytes with Robotic Experimentation and Machine Learning. Cell Reports Physical Science, 1(12).
[7] Moses, I. A., Joshi, R. P., Ozdemir, B., Kumar, N., Eickholt, J., & Barone, V. (2021). Machine Learning Screening of Metal-Ion Battery Electrode Materials. ACS Applied Materials and Interfaces.
