Developing a Machine Learning Model to Discover New Dielectric Materials

As technology advances and society pivots to more renewable sources of energy, the need for a better way to store and deliver power is increasing. Capacitors have high power densities, but limited total energy storage capacity. It would fill a growing need if we could improve their energy densities, but this will require superior dielectrics. Experimentally looking for new dielectric materials is a lengthy process because of the many possible combinations of composition and crystallographic structure. In this project, we use information about known materials to train a machine learning model to predict the dielectric constant and the dielectric strength of previously unstudied materials. We approach this problem in multiple steps: first, two classifiers categorize our unknown materials as metallic or insulating and stable or unstable. Second, two regressors predict the dielectric constant and band gap (as a proxy for dielectric strength), using primarily compositional information. Our best classifiers use a random forest model and we are able to predict both stability and metallicity with reasonable accuracy. The regressors are built using a voting model which successfully predicts band gaps but struggles to correctly identify dielectric constant. We expect that incorporating accurate dielectric constant predictions will likely require more thorough handling of crystallographic structure in the list of features. Nevertheless, we are able to generate a list of promising materials that helps narrow the search space for the next generation of dielectric materials.