In materials science, candidates for new functional materials are typically explored by trial and error through calculations, synthetic methods, and material analysis. However, the approach takes time and requires expertise. Now, Japanese researchers have used a data-driven approach to automate the process of predicting new magnetic materials. By combining first-principles calculations, Bayesian optimization, and monatomic alternative deposition, the proposed method can enable faster development of next-generation electronic devices.
Materials scientists are constantly looking for new “functional materials” with favorable properties geared towards certain applications. For example, the discovery of new functional magnetic materials could open the doors to energy-efficient spintronic devices. In recent years, the development of spintronics devices like magnetoresistive random access memory – an electronic device in which a single magnetoresistive element is embedded as a bit of information – has progressed rapidly, for which magnetic materials with high magnetocrystalline anisotropy (MCA) are required. Ferromagnetic materials, which retain their magnetization without an external magnetic field, are therefore of particular interest as data storage systems. For instance, L10-type ordered alloys consisting of two elements and two periods, such as L10-FeCo and L10-FeNi, have been actively investigated as promising candidates for next-generation functional magnetic materials. However, the combination of constituent elements is extremely limited, and materials with extensive element type, number, and periodicity have rarely been explored.
What hinders this exploration? Scientists point to combinatorial explosions that can easily occur in multilayer films, requiring a lot of time and effort in component selection and material fabrication, as the main reason. Furthermore, it is extremely difficult to predict the function of MCA due to the complex interplay of various parameters including crystal structure, magnetic moment, and electronic state, and the conventional protocol relies heavily on trial and error. errors. Thus, there are many possibilities and needs to develop an efficient way to discover new high performance magnetic materials.
On this front, a team of Japanese researchers including Prof. Masato Kotsugi, Mr. Daigo Furuya and Mr. Takuya Miyashita from Tokyo University of Science (TUS), as well as Dr. Yoshio Miura from the National Institute of Science materials (NIMS), has now turned to a data-driven approach to automate the prediction and synthesis of new magnetic materials. In a new study, which went online on June 30, 2022 and published in Science and Technology of Advanced Materials: Methods on July 1, 2022, the team reported their success in developing an exploration system of materials by integrating computation, information, and experimental sciences for high-MCA magnetic materials. Professor Kotsugi explains: “We focused on artificial intelligence and combined it with computational and experimental science to develop an efficient method of material synthesis. Materials that show promise beyond human expectation have been discovered in terms of of electronic structure. So it will change the nature of materials engineering!”
In their study, which was the result of joint research by TUS and NIMS and supported by JST-CREST, the team calculated the MCA energy by first-principles calculations (a method used to calculate electronic states and physical properties of materials based on the laws of quantum mechanics) and performed Bayesian optimization to search for high-energy MCA materials. After examining the Bayesian optimization algorithm, they found promising materials that were five times more efficient than by the conventional trial-and-error approach. This robust materials search method was less sensitive to the influences of irregular factors such as outliers and noise and allowed the team to select the top three candidate materials — (Fe/Cu/Fe/Cu), ( Fe/Cu/Co/Cu), and (Fe/Co/Fe/Ni) – including iron (Fe), cobalt (Co), nickel (Ni) and copper (Cu).
The top three predicted materials with the highest MCA energy values were then fabricated via the monatomic alternate stacking method using the laser pulsed deposition technique to create multilayer magnetic materials composed of 52 layers, namely [Fe/Cu/Fe/Cu]13, [Fe/Cu/Co/Cu]13and [Fe/Co/Fe/Ni]13. Of the three structures, [Fe/Co/Fe/Ni]1 showed an MCA value (3.74 × 106 erg/cc) much higher than that of L10-FeNi (1.30 × 106 erg/cc).
Additionally, using the second-order perturbation method, the team found that MCA is generated in the electronic state, which was not realized in previously reported materials. This attests to the relevance of using Bayesian optimization to identify electronic states that are likely impossible to imagine by human experience and intuition alone. Thus, the developed method can autonomously search for suitable elements to design functional magnetic materials. “This technique is extendable to advanced magnetic materials with more complicated electronic correlations, such as Heusler alloys and spin-thermoelectric materials,” observes Professor Kotsugi.