Researchers From MIT Have Developed A New Machine Learning Based Approach With 90 Percent Accuracy To Screen Candidate Materials If They Are Topological For Next-Generation Computer Chips or Quantum Devices

Topological materials are a special kind of material that have different functional properties on their surfaces than on their interiors. One of these properties is electrical. These materials have the potential to make electronic and optical devices much more efficient or serve as key components of quantum computers. But recent theories and calculations have shown that there can be thousands of compounds that have topological properties, and testing all of them to determine their topological properties through experiments will take years of work and analysis. Hence, there is a dire need for faster methods to test and study topological materials.

A team of researchers from MIT, Harvard University, Princeton University, and Argonne National Laboratory proposed a new approach that is faster at screening the candidate materials and can predict with more than 90 percent accuracy whether a material is topological or not.
The traditional way of solving this problem is quite complicated and can be explained as follows: Firstly, a method called density functional theory is used to perform initial calculations, which are then followed by complex experiments that involve cutting a piece of material to atomic-level flatness and probing it with instruments under high vacuum.

The new proposed method is based on how the material absorbs X-rays, which is different from the old methods, which were based on photoemissions or tunneling electrons. There are certain significant advantages to using X-ray absorption data, which can be listed as follows: Firstly, there is no requirement for expensive lab apparatus. X-ray absorption spectrometers are used, which are readily available and can work in a typical environment, hence the low cost of setting up an experiment. Secondly, such measurements have already been done in chemistry and biology for other applications, so the data is already available for numerous materials.

Now, after collecting data, the next step is finding the relationship between spectral data characteristics and topological properties. Since the spectral data is high-dimensional, a dense neural network is utilized to learn this relationship. Two types of information are being used by the model: first, what type of atom is being analyzed, and later, its X-ray absorption spectrogram.

Some fully connected layers operate on the atom-type label and spectral input to produce atom-type and spectral embedding. Now the spectra embeddings are assigned to element-specific channels through a direct product with the corresponding atom-type embedding. These features are then added together and passed into another network of Fully connected layers that predict a binary topological class.

Even though the model has more than 90 percent accuracy, as it is a deep learning model, the researchers don’t exactly know what is going on inside the box or why the model works. With future research prospects in mind, the team utilized the trained model to create a periodic table that shows the model’s overall accuracy on compounds made from each element. It will serve as a tool to help the researchers narrow down the compounds that may have topological properties. The current results demonstrate a promising pathway to use machine-learning-aided methods for screening candidates for topological properties. Thus, an X-ray absorption spectrogram with machine learning is a powerful tool to use in similar applications.

This Article is written as a research summary article by Marktechpost Staff based on the research paper 'Machine-Learning Spectral Indicators of Topology'. All Credit For This Research Goes To Researchers on This Project. Check out the paper and reference article.

Please Don't Forget To Join Our ML Subreddit