4-14、Descriptors in Solid-State Materials Informatics

Seiji Kajita

Program Manager, Materials Informatics Research-Domain

Toyota Central R&D Labs., Inc.

摘要：Materials informatics has been emerging as a high way for lean developments of advanced materials, driven by an expectation to shorten the period by a factor of 1/10. A virtual screening is the common strategy in which prospective compounds are selected from a vast chemical space beyond human capability without relying on experiments. The main engine of the virtual screening is machine learning, because it can execute rapid inferences of target properties.

However, this machine learning-based screening remains a fundamental challenge, which is so to say ”material data is not big data”. Given the fact that an image recognition task typically requires more than thousands of the training images, number of well-curated experimental data of materials is very limited. To predict a property by using a small amount of the supervised data, a key point is how to describe a material entity from a multifaceted perspective. A material consists of a set of ions and electrons which mutually interact to exercise several properties, which are encoded by descriptors. Involving more attributions by the ensemble descriptor is essential to complement the small data by a framework of ensemble learning.

Here I present solid-state descriptors to highlight this ontology issue in the material informatics, introducing a state-of-art descriptor and our original one [1] to represent a configuration of the ions and field quantities (e.g., electron density). Then, I show an example of a material search through a virtual screening by using the ensemble-scope descriptor.

参考文献：

[1] Kajita, Seiji, et al. "A Universal 3D Voxel Descriptor for Solid-State Material Informatics with Deep Convolutional Neural Networks." Scientific reports 7.1 (2017): 16991.

DOI:10.12110/secondfmge.20181014.414