A team of researchers from the Massachusetts Institute of Technology (MIT) utilizes artificial intelligence to accelerate the discovery of an intriguing phenomenon in special materials, a discovery that could lead to the development of electronics that operate without energy loss.
[תרגום מאת ד”ר משה נחמני]
Superconductors have long been considered the main approach to electron transfer without resistance. In the last decade, a new family of ‘topological materials’ has been developed that provides a promising alternative to the development of electronics that operate without energy loss. Compared to superconductors, topological materials have a number of advantages, such as resistance to interference. In the field of magnetism, the phenomenon of “magnetic proximity effect” is known, which occurs when the magnetism penetrates slightly into the surface of a topological material. At the same time, watching this result has so far been particularly challenging. Explains researcher Zhantao Chen from the Massachusetts Institute of Technology (MIT: “The presence of this result is usually too weak to measure using existing methods.”
In recent years, researchers have relied on measuring this result for a method known as “polarized neutron reflectometry” that allows testing of the deep magnetic structure of multilayer materials. In this method, two polarized neutron beams with an opposite spin are returned from the sample and reach the detector. “If the neutron hits a magnetic flux with an opposite direction, like the one inside a magnetic material, it will change its spin state, which will lead to measuring different signals from the two neutron beams with the spin up / down,” explains one of the researchers. As a result, the proximity result can be measured if a thin layer of essentially non-magnetic material – located next to the magnetic material – becomes magnetic. However, the result is very subtle, reaching a level of only one nanometer in depth, and its measurement becomes challenging and unequivocal when disturbed by the results of the experiment itself. “By integrating ‘machine learning’ into our method, we hope to get a clear picture of the mechanism,” explains the lead researcher. This hope was indeed realized when the findings of the research team were recently published in the prestigious scientific journal Applied Physics Review.
The researchers examined a topological insulator – a material that insulates electricity in its inner part, but is capable of conducting an electric current on its outer surface. They chose to focus on layered materials containing the topological insulator Bismuth Selenide (Bi2See3) Combined with the ferromagnetic insulator Europium Sulfur (EuS). Bismuth selenide, in essence, is a non-magnetic material, so the sulfur europium layer is responsible for the difference between the signals measured by the two polarized neutron beams. In addition, with the help of machine learning, the researchers were able to identify and quantify another contribution to the PNR signal: the induced magnetism created at the interface between the two substances bismuth selenide / europium sulfate. “Machine learning methods are very effective in identifying the basis of complex data, and make it possible to distinguish and differentiate between subtle results such as proximity magnetism in PNR measurements,” the researcher explains.
When the signal is first fed into a machine learning model it is very complex. The model is able to simplify this signal so that the proximity result is increased and becomes much more noticeable. Next, the model is able to quantify the induced magnetism – while determining whether the result of the magnetic proximity is indeed observable – along with other properties of the material system, such as: thickness, density and hardness of the various layers. “We have reduced the ambiguity obtained in previous tests, due to the doubling of the separation capacity resulting from the use of machine learning,” the researcher explains. The results mean that the researchers were able to distinguish between the properties of the materials at a length of 0.5 nanometers, about half of the typical spatial dimension of the proximity result. This situation is equivalent to a person watching the writing of a writing board about six meters away from him so that he cannot discern the separate words written there. However, if we reduce this distance by half, we may be able to read the separate words. The data analysis process can also be significantly accelerated by relying on machine learning. “In days gone by, the researchers were required to spend entire weeks analyzing all the data until an imaging curve was obtained that would match the experimental results curve,” the lead researcher explains. “A large amount of experimentation is required because the signal itself can match different combinations of metrics. The neural network provides you with an immediate answer. No more guesswork – no more experimentation and error.”
External factors praised the new study – which is the first of its kind to measure the effectiveness of machine learning in identifying proximity outcomes, and among the first in general to analyze PNR data. “The study provides an alternative way of identifying the meticulous details in PNR data, and shows how higher resolution is consistently achievable,” explains one of the professors in the field.
The research group is already planning to expand the scope of its experiments. “The result of magnetic proximity is not the only weak result that interests researchers today,” the researcher explains. “The machine learning infrastructure we have developed can be easily adapted to different types of problems, such as the proximity result of superconductors, which is one of the most important problems today in the field of quantum computing.”
The scientific article
MIT website review
More on the subject on the Knowledge website: