Our strategy is to build on the Electrolyte Genome concept by adding powerful machine learning techniques, including “deep” and “active” learning methodologies. Data will be obtained from state-of-the-art electronic structure calculations enabled by high performance computing. We envision that this powerful approach will minimize the number of experiments and calculations needed for addressing the science gaps.

The goal is to identify candidate molecules for energy storage materials (i.e., electrolytes to support flow batteries and polyvalent ion batteries), which require simultaneously optimizing various properties. These properties include redox activity, stability, viscosity, conductivity, solubility, and protective interphases for electrode materials.

Considering the enormous space of molecules, directly computing all properties of billions of molecules is impractical. Consequently, we will use a combination of computational simulation of smaller subsets of molecular properties along with experimental data, deep learning, and active learning to identify molecules with desired properties using the fewest number of calculations. This approach will enable us to navigate the enormous space of materials efficiently and accurately.

We will use an active learning framework that iteratively selects the next simulations, performs those simulations, and then integrates the results of those simulations into a deep learning model that evaluates where further simulations are required. The result of this work will be an unprecedented database of candidate materials much larger than databases available today, characterized in far more detail, and selected more intelligently that can identify new electrolyte molecules for synthesis and testing.