Data Preprocessing and Machine Learning Modelling for Battery Electrolyte Discovery

Abstract

Developing advanced battery technologies is vital to meet the energy, safety, and cost-efficiency demands of modern devices. The choice of electrolyte significantly impacts a battery’s energy output, charge-discharge rates, and safety, especially in high-power applications such as electric vehicles. Selecting the ideal electrolyte is challenging due to numerous solvent and salt combinations. In this study, machine learning techniques, including Chemprop, Extreme Gradient Boosting (XGB), and Kernel Ridge Regression (KRR), were employed to create predictive models for ionic conductivity in solvent-salt mixtures based on compositions, thereby narrowing the search space for discovering new electrolytes. The Chemprop model excelled better on test set, with R<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> value of 0.47, surpassing XGB and KRR, with R<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> values of 0.04 and -0.14, respectively. Furthermore, the cluster-split strategy outperformed the random-split strategy when tested on an out-of-distribution dataset, with R<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> values of 0.37 and -0.34 respectively. The results of this study show that leveraging the generalizable Chemprop model, especially with a cluster-split strategy, expedites electrolyte discovery, potentially advancing energy storage technology for cleaner energy use. It can be concluded that the Chemprop model is a more suitable model for predicting electrolyte properties, and the cluster-split strategy which increases the dissimilarity between the data sets is a more suitable data splitting technique for battery electrolytes data preprocessing.