Experimental design for the highly accurate prediction of material properties using descriptors obtained by measurement

Abstract

In materials science, both controllable and uncontrollable descriptors can be used to characterize materials. Examples of controllable descriptors include the composition of elements and fabrication processes; in contrast, uncontrollable descriptors are generated by experimental data characterizing particular samples, such as raw spectral data or specific gravity. In this study, we consider an experimental design to obtain a highly accurate prediction model where the uncontrollable descriptors of materials are features and its material properties are labels. In general, as uncontrollable descriptors are more closely related to material properties, predictions based on them will be more accurate. The goal of the experimental design in the present study is not the improvement of the material properties as such but the prediction of their properties. To realize this design, we select appropriate controllable descriptors for the synthesis of the candidate material that improve the prediction accuracy when the corresponding uncontrollable descriptors and material properties are added to the training data. We propose two experimental design methods, one based on Bayesian optimization and the other on uncertainty sampling. Using a polymer database in which controllable and uncontrollable descriptors, and mechanical properties are recorded, we confirm that our method can select an appropriate candidate material to train a highly accurate prediction model in which the material properties are predicted by uncontrollable descriptors. Our proposed method can be applied to materials developments where uncontrollable descriptors are more easily obtained by experiments than obtaining target material properties; it will also be useful for extracting the relationship between structure and properties of a material.

Keywords

• uncontrollable descriptors
• experimental design
• polymer