Discovery of Novel Dielectric Materials With Large Energy Bandgaps Using Statistical Optimization

We present results from a feedback-based materials design work-flow to find novel materials with optimized dielectric properties. The objective is to improve the performance of electrical devices that depend on charge bearing capacity, which directly depends on dielectric constants and band-gap energies of the compounds. A data-set containing data of dielectric tensors for 1864 materials was extracted from MaterialsProject.org¹ to train the statistical deep-learning models. The set of thermodynamically stable materials from OQMD.org² was used as the search-space to discover novel materials with large values for dielectric constants and bandgap energy. A reliable neural network model was built over the small training data and combined it with statistical optimization strategies. Dielectric properties of predicted materials were computed using Density Functional Perturbation Theory and fed back into subsequent generations of neural network models. Each design cycle in this approach successfully picked up new promising materials from the large search-space, including mixed anion compounds with very large bandgap and dielectric constants - a highly optimized scenario for industrial applications.
[1]Jain et al.,APL Materials,2013
[2]Saal et al. JOM,2013