Machine Learning for FB Electrolyte Screening

Experimental techniques and complementary simulations are constantly producing new information. Machine-learning (ML) techniques allow us to turn sets of independent data points into insights: based on a sufficient body of data that parameterize a given phenomenon, data-driven models can be trained to detect within the training set the underlying patterns and on this basis they can make predictions for new instances. This overview briefly discusses the relevant aspects and the critical issues in this process, starting from the data as the essential ingredients, some ways of representing said data as immediate input to an algorithm while preserving in some sense their original meaning, and various ML techniques that process these data to come up with trained models. The use and limits of ML techniques for flow battery (FB) research are illustrated by the standard potential and aqueous solubility. Although inherently limited by the scope of the training data set, ML models offer extremely efficient evaluation in e.g. high throughput screening jobs, allowing experimental and/or physics-based simulations to focus on the innovative research only.