MIT Researchers Discovered Hidden Magnetic Properties In Multi-Layered Electronic Material By Analyzing Polarized Neutrons Using Machine Learning

This research summary article is based on the paper from MIT  'Elucidating proximity magnetism through polarized neutron reflectometry and machine learning'

Please don't forget to join our ML Subreddit

Superconductors have long been thought to be the primary method for creating electronics without resistance. In the last decade, a new class of quantum materials known as “topological materials” has emerged as an alternate yet the intriguing approach to producing electronics with no energy dissipation (or loss). Topological materials have a few benefits over superconductors, such as shock resistance. The ” magnetic proximity effect is a critical approach to dissipationless electronic states is the “magnetic proximity effect.” This happens when magnetism penetrates slightly through the surface of a topological material. Observing the proximity effect, on the other hand, has been difficult.

The difficulty is that the signal people are seeking that would indicate the presence of this impact is frequently too weak to detect decisively with existing methods.

Source: https://aip.scitation.org/doi/pdf/10.1063/5.0078814

What’s underneath and between the layers?

For the past several years, researchers have relied on polarised neutron reflectometry (PNR) to investigate the depth-dependent magnetic structure of multilayered materials and to hunt for phenomena such as the magnetic proximity effect.

PNR involves the reflection of two polarised neutron beams with opposing spins from a sample and their collection on a detector. Suppose the neutron encounters a magnetic flux with the opposite orientation found inside a magnetic material. In that case, it will change its spin state, resulting in different signals measured from the spin up and spin down neutron beams. Hence, the proximity effect can be recognized if a tiny layer of a typically nonmagnetic substance gets magnetized when placed closely near a magnetic material.

However, ambiguities and difficulty interpreting experimental data might develop because the impact is modest, spanning only approximately 1 nm in depth.

The researchers looked at topological insulators, which are materials that are electrically insulating on the inside but may carry electric current on the outside. The layered materials system comprises the topological insulator ‘bismuth selenide’ (Bi2Se3) and the ferromagnetic insulator ‘europium sulfide’ (EuS).

Because Bi2Se3 is a nonmagnetic material in and of itself, the magnetic EuS layer dominates the difference in signals obtained by the two polarised neutron beams. However, the researchers could detect and quantify another component of the PNR signal using machine learning. This was the magnetization produced in the Bi2Se3 at the interface with the neighboring EuS layer. Machine learning approaches are particularly successful in extracting underlying patterns from complex data, allowing us to distinguish minor effects like proximity magnetism in the PNR measurement.

The PNR signal is quite complicated when it is first input into the machine learning model. The model can simplify this signal such that the proximity effect is increased and hence more noticeable. The model can then determine the induced magnetization — as if the magnetic proximity effect is observed — and other properties of the materials system, such as the constituent layers’ thickness, density, and roughness, using this simplified representation of the PNR signal.

The uncertainty in prior studies has been minimized due to the doubling in resolution gained using the machine learning-assisted technique. Material qualities were detected at length scales of 0.5 nm, which is half the average spatial range of the proximity effect. That’s like gazing at a whiteboard from 20 feet away and not being able to see. But if the distance is cut in half, you might be able to read it all.

Using machine learning to analyze data can also drastically speed up the process. In the previous days, weeks of tweaking with all the settings were required before the simulated curve met the experimental curve. It may require several tries because the same [PNR] signal may correlate to multiple parameter combinations.

The neural network responds immediately. There’s no more speculating or trial and error. As a result, the framework has been placed in a few reflectometry beamlines to aid in examining a more extensive range of materials.

Some outside observers have commended the new study, which is the first to assess the usefulness of machine learning in detecting the proximity effect and one of the first to apply machine-learning-based programs for PNR data processing. 

The MIT-led team is already thinking about broadening the scope of their research. The machine learning framework is easily adaptable to various challenges, such as the superconducting proximity effect, which is of significant interest in quantum computing.

Paper: https://aip.scitation.org/doi/pdf/10.1063/5.0078814

Reference: https://news.mit.edu/2022/seeing-elusive-magnetic-effect-through-lens-machine-learning-0324