Abstract: The vast space of molecular materials often renders high-throughput experimental and computational screening efforts intractable. To address this challenge, a typical approach is to train a machine learning model on known materials and then make predictions for the unknown search space. However, whether such approach is reliable depends on our confidence in a sufficient large and diverse training set, a factor that is not easily quantified. On the other hand, an active learning model yields uncertainty quantification for every prediction. Therefore, promising materials can be selectively evaluated and added to the training set with confidence, regardless of the size of the search space. 

In this talk, the basic components and working principles of an active learning scheme based on Bayesian optimization (BO) is introduced. Then, the development, implementation, and evaluation of a BO model for energy storage application will be discussed.