Materials Informatics with Limited Data

Data-driven science is advancing rapidly and transforming the world. As such, machine learning has received considerable attention as a key driver of the next frontier of materials science, substantially reducing the time and costs required in the discovery and development of innovative materials. Material design aims to identify a set of design variables that exhibit the desired properties. To address this, we apply a two-stage workflow consisting of forward and backward predictions. The objective of the forward problem is to predict the properties of given design variables. The inverse problem involves identifying promising design candidates that exhibit a given set of desired properties by solving the inverse mapping problem of the forward predictive model. This chapter describes the basic concepts and key technologies of machine learning for the forward and inverse analyses of material design, including Bayesian molecular design, transfer learning, functional output regression, and automated molecular dynamics simulations for polymeric materials, illustrated with practical applications from our recent work on polymer and quasicrystal research. These general techniques are also applicable to drug molecule design.