Quasicrystals discovered using machine learning algorithm 'for the first time in 40 years' — Technology for predicting thermally stable chemical compositions developed by ISM

A joint research group led by Professor and Director Ryo Yoshida of the Data Science Center for Creative Design and Manufacturing at the Institute of Statistical Mathematics (ISM), together with the Tokyo University of Science and the University of Tokyo has announced the development of a machine learning technique for predicting chemical compositions that form thermally stable quasicrystals by reading patterns of quasicrystals and related materials that have been synthesized previously. Based on this machine learning prediction, three new quasicrystals (Al₆₅Ni₂₀Os₁₅, Al₇₈Ir₁₇Mn₅, and Al₇₈Ir₁₇Fe₅) were discovered. These are the first quasicrystals ever discovered by machine learning algorithms in the approximately 40-year history of quasicrystal research. The results were published on September 25 in the international journal Physical Review Materials.

It is said that approximately 90% of the quasicrystals found on Earth today have been derived from a series of materials discovered by Dr. An-Pang Tsai, a former Professor of the Institute of Multidisciplinary Research for Advanced Materials (IMRAM) at Tohoku University, and his colleagues.

The discovery of new quasicrystals has led to the discovery of new physical phenomena such as anomalous electronic properties, insulator-like behavior, valence fluctuations, quantum criticality, superconductivity, and ferromagnetism. However, the mechanisms behind the formation and stabilization of quasicrystals are not well understood, and guidelines for designing new materials in this context have not been established. Consequently, research on quasicrystals has not progressed.

Therefore, the research group introduced machine learning to accelerate the process of discovering quasicrystals. The prediction model takes the chemical composition as input and outputs a class label indicating whether the material would form quasicrystals or not. The training data consists of the chemical compositions of quasicrystals, related materials, and ordinary periodic crystals that have been synthesized to date.

The results demonstrate that this model can perform the binary classification task of predicting whether a quasicrystal would be formed or not with an accuracy of over 95%. It was also found that this machine learning algorithm learns empirical rules from the patterns of quasicrystal composition discovered by Tsai and his colleagues. Furthermore, the five rules of quasicrystal formation were revealed by extracting the input-output rules from the black box model of machine learning.

The formation rules can be expressed in five simple mathematical expressions for constraints such as the van der Waals radius of an atom (equal to half the distance between them if the electrostatic forces between two unbound atoms are balanced), electronegativity (the strength of the force with which an atom attracts electrons), and others.

Using these rules, an exhaustive screening was conducted on 1080 alloy systems, corresponding to the entire space of aluminum ternary alloys. Finally, the number of candidates was narrowed down to 30, and synthesis experiments were conducted by selecting Al-Ni-Os, Al-Ir-Mn, and Al-Ir-Fe as the first attempts. The results revealed the presence of a quasicrystalline phase (Al₆₅Ni₂₀Os₁₅, Al₇₈Ir₁₇Mn₅, and Al₇₈Ir₁₇Fe₅) in all the systems. All three quasicrystals were observed after a long annealing process, indicating their thermodynamic stability.

Yoshida stated, "The adoption of materials informatics techniques has rapidly advanced in various fields of materials science in recent years. This discovery marks the first step towards data-driven quasicrystal research. Although about 40 years have passed since the discovery of quasicrystals, little is known about the formation conditions and stabilization mechanisms of these structures. This study has demonstrated that artificial intelligence can significantly contribute to unraveling the unresolved challenges in quasicrystal research. Moving forward, the goal is to discover novel materials, such as quasicrystals with semiconductor-like properties and antiferromagnetic quasicrystals, which are yet to be explored by humanity."

Journal Information
Publication: Physical Review Materials
Title: Quasicrystals predicted and discovered by machine learning
DOI: 10.1103/PhysRevMaterials.7.093805