Introduction of Research
Harmony between Materials Science and Information Science
Our approach explores new materials and functions by statistical learning and factors that control material properties utilizing data obtained by experiments and theoretical calculations. We call this "Materials Informatics", which has drawn much attention recently. Such development is a natural application of data science. However, around 2012, when we began planning this research, there were only few researchers around the world who could understand and practice both materials science and information science. Hence, we began this project from scratch. Therefore, we shared our research goals and technical terms during regular meetings with mainly young researchers in both fields and initiated a number of collaborative research projects.
The efforts have come to fruition as creative achievements. For example, we efficiently obtained equivalent information by statistical learning of a small number of datasets to realize effective data collection methods instead of employing costly first-principles calculations to cover the entire search space. This not only shortens the time to explore but also significantly extends the search space within a specified research time. Unlike conventional material research, where exploration is done after constructing an exhaustive database, we have achieved a decisive difference in efficiency.
Materials Science & Information Science Collaboration Cooperation Task Force
To achieve close collaborations between materials science and information science, researchers in both fields have formed a task force and communicate regularly. Thanks to sharing common research goals and technical terms, our integrated research has produced results that have exceeded our initial expectations.
Highly Efficient Discovery of New Materials and New Functions by a Virtual Screening Method
It is rare in materials science to have property data for the entire search space. Therefore, machine learning using a small data subset in the entire search space is useful in order to predict the properties of the remaining materials. This method is called "virtual screening", which not only shortens the time required for research and development but also greatly expands the search space for a given time period.
In this project, we have conducted a virtual screening on ultra-low thermal conductivity materials and have efficiently discovered many materials with thermal conductivities one digit or more lower compared to conventionally known low thermal conductivity materials. As the initial data, we used the result of anharmonic phonon first-principles calculations developed in-house, which enables the lattice thermal conductivity to be calculated with an accuracy comparable to experiments. However, it is not realistic to apply it to many materials due to its high computational cost.
In this project, we have efficiently discovered ultra-low thermal conductivity materials of a 0.1 W/mK level at 300 K from 55,000 known materials based on the calculation results of about 100 materials. This widely expands the selections in the development of new thermoelectric materials.
A. Seko et al. Phys. Rev. Lett. 115 (2015) 205901.
Exploring Complex Ion Conduction Paths by First-principles Calculations and Bayesian Optimization
One method to theoretically predict the ionic conductivity of a solid is potential energy surface (PES) mapping of conduction carriers in a crystal. This method introduces a fine grid in a host crystal and comprehensively calculates the potential energy (PE) of the carriers at each grid point. Its calculation cost increases sharply as the symmetry of the crystal decreases.
In this project, we have constructed a methodology to evaluate PES with high speed and high accuracy by combining first-principles calculations and Bayesian optimization. Specifically, we focus on the fact that the region dominating ion conduction is a part of the whole crystal and selectively evaluate the dominant region.
Using this method, we have evaluated PES of cubic perovskite-type BaZrO3 with proton conductivity and accurately estimated the potential barrier ΔE mig for long-distance conduction using only 30 out of the total 1,768 grid points. Furthermore, we have evaluated tetragonal Scheelite-type LaNbO4 with anisotropic conduction and estimated ΔE mig in the ab-plane and the c-axis direction using only 100 points.
Utilizing this PES evaluation method, it is possible to quickly and accurately conduct screenings for enormous multicomponent materials and to develop solid state ionic materials according to theoretical calculations.
K. Toyoura et al. Phys. Rev. B 93 (2016) 054112.
Development of Machine Learning Software to Support Material Scientists
Materials sciences often find research using machine learning challenging. For example, coding a machine learning algorithm from scratch would take an enormous amount of time and effort; the start of actual research may be considerably delayed. To address this, we have developed "COMBO", which is software of a Bayesian optimization algorithm indispensable for developing materials. COMBO was made it public in 2016. To make it easy for materials researchers to use, we devised it so that parameter tuning is unnecessary. In addition, we adopted an algorithm that can handle large amounts of data at high speeds.
COMBO is currently used by many researchers both in Japan and around the world. For example, it is utilized to find the optimum interface structure and to improve the efficiency more than 100 times compared to comprehensive exploration. For a larger exploration space, a Monte Carlo tree search (MCTS) adopted in the computer Go game is useful. The figure below is a conceptual diagram of software "MDTS" developed in this project to exploit MCTS for atomic arrangements of materials. For material development, diverse machine learning algorithms are required.
T. Ueno et al. Materials Discovery 4 (2016) 18.
