Inverse Design of Organic Mixtures Using First-Principles Calculations Mediated Bayesian Optimization

Various properties of organic mixtures used for electronic materials in smart device, memory module and IC packaging generally come from their constituents. There are several factors that affect the target properties of organic mixtures, among which the molecular combinations and their compositions are the most important. Traditionally, to find the optimal input for desired properties, a trial-and-error experiment method has been used, but it is costly because many samples need to be prepared and tested. Machine learning based optimizations can be employed to optimize compositions, but they are limited to handling fewer than 20 variables. In this study, we optimize molecular combinations and compositions of over 750 organic materials through first-principles calculations mediated Bayesian optimization. Because optimization directly from targets to vast number of compositions are difficult, we first encoded composition space of experimental dataset to physical property space using first-principles calculations. Then, these encoded data are used as inputs for optimize from targets with Bayesian optimization. We obtain optimal molecular combinations and their compositions from optimized physical property. Through this approach, we achieve average optimized errors (%) of less than 3%. This study holds considerable significance for effectively optimizing combinations of a wide range of organic materials to achieve desired products.