Machine Learning Creates New Technologies


HomeHome / Blog / Machine Learning Creates New Technologies

Apr 27, 2023

Machine Learning Creates New Technologies

Machine Learning to Learn New Technologies This week I was at the 2023 IEEE

Machine Learning to Learn New Technologies

This week I was at the 2023 IEEE Intermag Conference in Sendai, Japan. This is a conference put on by the IEEE Magnetics Society (my first IEEE Society, member for 45 years). I was invited to attend as the President Elect of the IEEE. There were over 1,700 total physical and virtual attendees with close to 1,500 people at the conference in person. I believe that this is the largest magnetic conference since the Covid pandemic began in 2020.

I attended a session that had papers on applications of artificial intelligence for magnetic materials research. This is an example of discussions going on in the scientific and engineering community on how people can effectively use new AI tools to accelerate and assist in our understanding of the physical world and its applications to real world applications. These include making better magnetic memory devices, more efficient motors and many other practical activities.

This session included Mingda Li, from MIT who said that "data-fitting is one among many other uses that can be benefited from machine learning. The other is the focus on exploring hidden data, or building structure-property relations." For this latter application, the papers in this session utilized large material data bases. Mingda mentions a 146,000 materials database in this paper.

Y. Iwasaki from the National Institute for Materials Science, Tsukuba, Ibaraki, Japan used an autonomous materials search system combining machine learning and ab initio calculation to find multi-elemental compositions that could find alloy magnetizations higher than Fe3Co (the material at the peak of the Slater-Pauling curve). The image below, shows the results of this materials search over a 9-week period, gradually finding ways to increase the intrinsic magnetization of the modeled alloy.

Multi-week simulation to increase material magnetization

This research indicated that adding a bit of Ir and a bit of Pt could increase the magnetization of an iron cobalt alloy. When some physical iron cobalt iridium and iron cobalt platinum allows were made and measured it was found that about 4% Ir did indeed increase the magnetization of the FeCo alloy. Likewise, a little Pt in an FeCo alloy also increased the magnetization. Although alloy compositions with magnetization higher than Fe3Co have been found before, this investigation showed an example of how AI could be used as a tool for new material discoveries.

Claudia Felser and colleagues, from the Max Planck Institute of Chemical Physics of Solids as well as Spain, the US and China talked about using AI methods to develop new materials for what are called topological magnetic materials. These exploit chiral electron states on the bulk, surfaces and edges of solid objects. In physics, a chiral phenomenon is one that is not identical to its mirror image. Electron spins impart a chirality to an electron. She showed how materials with very high anomalous Hall Effect and a large anomalous Nearst effect were identified. An interesting element of this work relates to the interaction of gravitation in light matter interactions with magnetic topological materials. Perhaps these phenomena could provide new ways to detect and understand gravitation?

Masafumi Shirai and associates from Tohoku University used a large database of magnetic properties for what are called Heusler alloys interacting with an MgO tunneling layer for magnetic tunnel junctions (MTJs). Using machine learning and this database they were able to predict the Curie temperature of four component alloys (the temperature at which the magnetization goes to zero) and what is called the exchange stiffness (the exchange stiffness represents the strength of what are called exchange interactions among neighboring magnetic spins) at the interface with the MgO. Note that MTJs are used as the read sensors in hard disk drive and magnetic tape heads and in commonly used magnetic sensors.

The last paper in this session given by Alexander Kovacs with co-authors from Austria and Japan talked about using machine learning combined with finite element analysis of permanent magnetic material crystal grains to create more efficient motors and use less rare earths, for instance, for windmills. They optimized the chemical composition and microstructure of the magnet using machine learning models developed through assimilation of data from experiments and simulations. They demonstrate how high-performance, Nd-lean magnets can be created using the machine learning methods.

Machine learning is finding increasing uses in developing new materials, including magnetic materials used for digital storage. Various approaches can be taken, but using databases of known materials these models can predict the properties of new materials, making and virtually evaluating combinations much faster than a human could. Although not infallible, these approaches can speed scientific and engineering discoveries.