Machine Learning on Lithium Sulfur Materials

Materials properties are collective phenomena in a nonlinear relationship of compositions and interatomic configurations, which are difficult to acquire from computations. Interactions between the constituent particles (atoms) are so complicated which only the limited scale of the problem is analytically solvable with descriptive theory (e.g., density functional theory, molecular dynamics, etc.) combining with current high-performance computing technology. To generalize materials simulation and properties prediction, data-driven approaches (e.g., machine learning, neural network, etc.) are much needed. Here, we provide an example showing how the cohesive energy, Fermi energy and band structures of lithium-sulfur materials as intrinsic collective behaviors are learned via Machine Learning methods, such as random forest, K-nearest neighbor algorithm, LASSO, Neural Networks, etc, and the performance comparison of the different methods will be demonstrated, which provides a general design guidance for lithium-sulfur batteries.