A neural network for microstructure analysis

Research engineer Ivan Polyakov from Tomsk Scientific Center of the Siberian Branch of the Russian Academy of Sciences is developing software for the automatic analysis of a large number of microstructure images obtained using metallographic and electron microscopes. With a user-friendly interface, a neural network created and trained specifically for the purposes, and a bunch of original algorithms, scientists hope to gain a streamlined access to material property data, such as porosity, grain size, and phase composition.

– Oftentimes, when analyzing images, such as when it’s necessary to distinguish between the boundaries of various particles, we end up using different photo editing tools. Analyzing a couple of images manually is fine, but what if they come in vast numbers! The objectivity of the results in that scenario is compromised to the extent that it becomes a myth. That’s how we came up with the idea at TSC to look for a software solution that would allow for tackling the task, – explains Ivan.

The young scientist has a significant workload ahead! The future software bundle will consist of a neural network, several algorithms, and a user interface. As Ivan Polyakov explained, it is essential to build a neural network from scratch and train it on a carefully selected set of real microstructure images. The more variations the neural network learns, the more accurately it will be able to recognize them in real samples later.

The software runs on several algorithms that analyze porosity, grain size, and phase composition of the material’s microstructure. An algorithm for determining porosity has already been written, helping scientists quickly gather a large array of statistical data. The most challenging so far has been the algorithm for the analysis of grain sizes: Ivan is searching for an effective way to explicitly “explain” to the program what grain boundaries are and how they differ from pores. In addition, a user-friendly interface is already in place that allows for handling the program seamlessly and intuitively.

– The software is already helping scientists at Tomsk Scientific Center to obtain important preliminary data on the structure of the studied samples, porosity, and the sizes of particles and grains in powdered mixtures of the original reactive components, as well as in the chemical compounds (intermetallics, metalloids, complex composite materials) synthesized from them. This project was recently presented at “Technoprom-2024”, – noted Oleg Lapshin, head of the Department of Structural Macrokinetics at Tomsk Scientific Center.

In the future, the software will no doubt be in demand in materials science institutes and research centers, as well as in production, where it will help effortlessly identify various defects in the microstructure of manufactured products, and the neural network should be able to pick up on specific tasks in no time.