NUST MISiS: Method for studying the composition of material using machine learning

Main article: Machine Learning

2023: Research method using machine learning to decipher radiographs

NUST MISIS On April 10, 2023, it was reported that the student described a method that would allow materials scientists to save time money at. to interpretation roentgenograms He proposed using machine learning data X-ray diffraction to predict the phases of the crystal structure of transition metals and their oxides.

One of the main methods used in materials science - X-ray phase analysis - is based on obtaining data on the chemical composition of the material using X-ray diffraction. In practice, in the manufacturing process and during diffraction, various oxides and excess compounds are formed in the diffractometer plant, which may interfere with the identification of material phases, so there is still a need for more reliable and accurate methods for determining elements in diffractograms.

Machine learning has long been used to predict material properties, analyze crystal structure, and classify diffraction patterns. The radiograph database allows identification in a more specific and targeted manner, reducing the likelihood of interpretation errors. However, earlier in scientific work, such approaches were practically not applied to transition metal oxides, which are used in various fields - from pipeline transport to electronic devices.

The method consists of three stages. First, signs of peaks on spectograms are collected, that is, the position, value, distance and area of ​ ​ each peak is calculated. It is worth paying attention to the main feature - the area of ​ ​ the graph between peaks. It allows you to compare diffraction patterns of different substances and find exact matches. This algorithm counts the number of matches. Since peaks can have different deviations, machine learning is used. The resulting features, including the number of matched peaks, are fed to classical machine learning algorithms to correct the final result. The best model was a random forest with increased depth, told the author of the study Maxim Zhdanov.

One of the key functions used for this analysis is the calculation of peak area, which is used to quantify the intensity of diffraction peaks. This aspect makes it possible to more accurately and quickly identify different phases in radiographs, which in the future can significantly save the scientist time when examining the material.

The described method has its limitations. Identification of close phases can still be inaccurate in cases where the pattern and crystal structure differ for a number of reasons. Among other things, the accuracy of identification is influenced by the state of the diffractometer and previous experiments conducted.

The proposed method is difficult to work with tasks where many different substances are found. It is worth trying the use of neural networks with a structured latent space, for example, as in variational autoencoders, to extract more important features from different groups of materials, noted Maxim Zhdanov.