The development of crystalline materials found in semiconductors, pharmaceuticals, photovoltaics and catalysts continues to surprise the scientific community. These materials, made up of atoms, ions or molecules arranged in an ordered three-dimensional structure, are at the heart of many innovations. Their development requires precise methods for their identification, especially if they are multiphase, i.e. contain different types of crystals.
Currently, X-ray diffraction of powder samples is the predominant technique for this identification. However, the complexity increases with multi-phase samples. There, the precise identification of the different phases largely depends on the expertise of the scientists, making the process long and arduous. To counteract this, innovative data-driven methods such as machine learning are increasingly being used.
In a major breakthrough, a research team led by Associate Professor Tsunetomo Yamada of the Tokyo University of Science, in collaboration with the National Defense Academy, the National Institute of Materials Science, Tohoku University and the Institute of Statistical Mathematics, proposed a new “binary” for machine learning before classifier model. This model, published on November 14, 2023 in the journal Advanced Science, is capable of identifying phases of icosahedral quasicrystals (i-QC) from multiphase powder X-ray diffraction patterns.
To develop this model, the researchers created a “binary classifier” using 80 types of convolutional neural networks. After the model was trained with synthetic diffraction patterns representing i-QC phases, its performance was evaluated using synthetic patterns and a database of real patterns.
Researchers propose a machine learning model to identify a new Al-Si-Ru-i QC phase. This model shows significant potential to accelerate the identification process of multiphase samples.
Photo credit: Tsunetomo Yamada/TUS
The model achieved a prediction accuracy of over 92% and was able to successfully identify an unknown i-QC phase in Al-Si-Ru alloys when screening 440 measured diffraction patterns of unknown materials in six diffraction systems. The presence of the unknown i-QC phase was confirmed by analyzing the microstructure and composition of the material using transmission electron microscopy.
This deep learning model can identify the i-QC phase even if it is not the most dominant component in the mix. In addition, it can be used to identify new decagonal and dodecagonal quasicrystals and can be extended to various types of other crystalline materials.
Dr. Yamada concludes: “Using the proposed model, we were able to accurately detect unknown quasicrystalline phases in multiphase samples. The accuracy of this deep learning model opens up the possibility of accelerating the phase identification process of these samples.”