(595h) Data-Driven Prediction of Materials Properties in an Automated Fashion

There is pressing need for the use of rapid and reliable data-driven prediction scheme for materials development and optimization. It can drastically speed up the process of assessing key control variables for materials properties, avoiding the needs of scanning the entire design space with costly experimental measurements and computationally intensive simulations. However, complexity of data generation, model building, and validation procedures for the learned-model approaches could pose as a major obstacle, making them less accessible from the materials science and engineering community. In this work, we showcase the latest development in computerized algorithms for automated generation and ranking of predictive regression models, which is readily available for the design of new chemistry in molecular space. The methodology is demonstrated with large-scale virtual screening of organometallic phosphors, identification of design space for thermally activated delayed fluorescence (TADF) materials, turnover frequency prediction for Ziegler-Natta catalysts, and design of organic electronic compounds with high glass transition temperature (T_g). The automated data-driven predictive scheme provides unbiased measures to quickly assess the key design rules for a wide variety of applications, which could significantly lower the barrier towards large-scale virtual screening for developing novel materials solutions.