Hokkaido University develops method for perovskite material development — From prediction to synthesis and evaluation in an integrated process

A research group led by Professor Keisuke Takahashi, Assistant Professor Lauren Takahashi, Postdoctoral Researcher Fernando Garcia-Escobar, First-year Doctoral Student Tomoya Tashiro, and Second-year Master's Student Kenshin Shibata from the Faculty of Science at Hokkaido University has established a method for developing perovskite inorganic materials that can precisely predict and design bandgaps (indicators of light absorption) using machine learning. Their work was published online in Chemical Science.

Previously, perovskite materials were known for their excellent structure capable of efficiently absorbing solar light, but material design was difficult because their bandgaps fluctuate significantly with slight structural changes.

The research group generated thousands of descriptors (features) from structural information and elemental characteristics based on bandgap information from 282 perovskite-type compounds collected from past experimental literature. Subsequently, they selected optimal descriptors using their proprietary MonteCat method algorithm and constructed a Support-Vector Regression (SVR) model. This enabled them to develop a model that predicts bandgaps with high precision based on structure and composition.

Next, using this model, they predicted the bandgaps of 1,852 hypothetical perovskite compounds theoretically considered stable. From these, they selected 86 promising candidates that satisfied both bandgaps suitable for solar light applications (0.45 to 2.2 eV) and structural stability. Of these, they actually synthesized four compounds (LaCrO₃, LaFeO₃, YCrO₃, YFeO₃) and confirmed through X-ray diffraction (XRD), scanning electron microscopy (SEM-EDS), and ultraviolet-visible spectroscopy (UV-vis-NIR) that they possessed the predicted characteristics for structure, composition, and bandgap. Through this research, the group successfully constructed a new material development workflow that integrates machine learning and experiments: extracting descriptors from structural and compositional information → predicting bandgaps → actual synthesis → verification through property evaluation. Research that achieves such high prediction accuracy using a descriptor-based approach for the highly sensitive property of bandgaps, and further matches synthesis results, is extremely rare. This is expected to become a standard model for future material development.

These results will not only contribute to accelerated development of energy-related materials such as solar cells, photocatalysts, and water splitting catalysts, but also strongly promote informatics-driven synthesizable material design starting from property design. In the future, the discovery of more diverse and highly functional materials is expected through the application of descriptor-based approaches to other property domains such as magnetism, dielectric properties, and thermoelectrics.

Journal Information
Publication: Chemical Science
Title: Designing and synthesizing perovskites with targeted bandgaps via tailored descriptors
DOI: 10.1039/D5SC04813C