From High-Throughput Experimentation to Data-Driven Analysis in the Development of Liquid Electrolytes for Lithium-Ion Batteries

Abstract

Lithium-ion batteries (LIBs) are considered one of the best candidates to meet the skyrocketing demand for high energy density energy storage systems that was sparked by increasing popularity of portable electronics, electro-mobility and the necessary shift towards renewable energy sources.1 While the demand for energy storage systems rises rapidly, research moves forward rather sluggishly due to the complex interplay between the various active and inactive materials of a battery. One of the bottlenecks is the large time requirement and limited knowledge gain of typical, sequential experiments where one experiment is performed after the other. To mitigate this limitation, high-throughput experimentation (HTE) setups have been designed, where multiple experiments are performed in parallel rather than in sequence. Although they are common in other fields of research, high-throughput methods have not been widely adopted in battery research yet.2 Nevertheless, once a HTE setup is established, the bottleneck tends to move from the experiment to the data processing and analysis. Suitable tools to circumvent this are modern machine learning (ML) techniques, recognized as invaluable tools for accelerated material development. In most studies conducted for LIBs, only a limited number of electrolyte formulations is considered for a given set of individual electrolyte components due to aforementioned limitations of experimental capabilities. In this work, we have employed a unique HTE setup that is capable of applying electrochemical impedance spectroscopy (EIS) to 96 liquid electrolytes in a broad temperature range in parallel to study one of the fundamental bulk properties of liquid electrolytes; the ionic conductivity.Two large datasets of ionic conductivities were acquired for electrolytes consisting of lithium hexafluorophosphate (LiPF6), ethylene carbonate (EC), ethyl methyl carbonate (EMC) and either vinylene carbonate (VC) or propylene carbonate (PC) in the temperature range of -30 to 60 °C. Conducting salt concentration as well as solvent ratios were varied in wide ranges and experimental results bundled with all necessary metadata in an in-house developed .json file format for each individual experiment. More than 5000 individual conductivity values have been determined for 128 different electrolyte formulations in total. The conductivity datasets were used by our project partners in different ML approaches. Linear and Gaussian process regression were used to unravel the impact of individual electrolyte components on the ionic conductivity.3 Symbolic regression was used to discover an equation that directly models the conductivity based on the electrolyte formulation and has some common features with the empirical Debye-Hückel-Onsager equation.4 Active learning was used to improve an initial predictive model with as little additional experiments as possible in order to generate an accurate model for a given dataset.5 The presented work serves as a proof of concept in a system where the relation between parameters and observable are comparatively simple and is the foundation for applying similar principles of data processing, metadata handling and ML application to more complex experiments like galvanostatic cycling.