An Active Learning Approach for the Design of Doped LLZO Ceramic Garnets for Battery Applications

Growing demand in applications like portable electronics and electric vehicles calls for cost-effective, safe, and high-performance energy storage systems. Development of solid-state electrolytes with Li $$⁺$$ ionic conductivities comparable to those of the current liquid chemistries is an important step towards meeting these needs. Unfortunately, one of the most promising solid electrolytes known to date, lithium lanthanum zirconium oxide (LLZO) garnets, exhibits far from ideal ionic conductivity. Thus, significant efforts, often through aliovalent substitution, have been devoted to increasing their ionic conductivity. Given the high-dimensional design space involved and the time required for synthesis, processing, and characterization of new materials, brute force approaches are not ideal to identify optimal compositions. We assess whether machine learning tools can be used to effectively explore the design space of LLZO garnets and potentially reduce the number of experiments involved in their development. We collected, curated, and filtered all the experimental results of Li $$⁺$$ ionic conductivity in LLZOs published in the scientific literature. Exploration of this data provides insights into the mechanisms that govern ionic transport in these oxides. Furthermore, we show that active learning with predictive models based on random forests can effectively be used with current data for the design of experiments. Our results indicate that the current highest Li $$⁺$$ ionic conductivity garnet LLZO could have been discovered with only 30% of the experimental studies conducted to date. All data and models are available online and can be used to drive future investigations.