Lectures

Neural networks and neutron tomography

The popularity of neural networks has been growing in recent years. Their idea is to simulate as closely as possible the functioning of the human brain: namely, its ability to learn and correct errors. This is one of the areas in the development of artificial intelligence systems. Using neural networks, one can generate texts, edit music, create images and write code, for example. And it can be successfully used in science. On 16 May, during the online lecture "Neural networks for physics", FLNP junior researcher Bulat Bakirov spoke about the challenges JINR scientists face and how they use neural networks in practice.

So, what is a neural network and how does it work? "In our understanding, a neuron is a mathematical model, the input of which is supplied with some data, they are summed up, nonlinear functions are applied and some result is obtained," Bulat Bakirov says. "Single neurons are united into layers and then arranged together. The neural network needs to be trained. And here everything is not so simple. The magnitude of the error is initially very large. And it is necessary to rearrange the parameters so that it decreases. Such a restructuring of network parameters is called one unit training epoch. After a certain number of training epochs, we can obtain a result that satisfies us."

This is a regular neural network, where each neuron is united to each next and subsequent layer. In fact, it works great for a number of tasks, but not those that involve images. More complex issues are met by the so-called convolutional neural network. This special architecture of artificial neural networks obtained its name due to the occurrence of a convolution operation, the essence of which is that each image fragment is multiplied by the convolution matrix (kernel) element by element and the result is summed up and written to a similar position in the output image.

How can convolutional neural networks be used, for example, in science? At FLNP, scientists use them in neutron tomography tasks, as Bulat Bakirov says. To construct images, you can use not only visible light, but also ionizing radiation. In this case, you can obtain information not only about the external, but also about the internal structure of the object. Compared to X-rays that are of an electromagnetic nature, neutron radiation is characterized by deeper penetration into the depths of the object under investigation that allows it to be used to study fairly large objects and to address a wide range of interdisciplinary scientific issues.

At FLNP, on beamline №14 of the high-flux pulsed reactor IBR-2, a specialized facility was developed for research using neutron radiography and tomography. This facility allowed to develop a new applied research area related to non-destructive analysis of the internal structure of a wide range of objects, products and materials, including archaeological sites of cultural heritage.

"Colleague archaeologists send us some samples and artifacts from excavations for research. We study them using the IBR-2 reactor and afterwards, we try to reconstruct the technology for making these artifacts, to study their internal structure and to determine the degree of their preservation," the scientist notes.

What results can be obtained in this way? For example, when examining medieval silver coins, you can see that some of them are counterfeits, made back in the Middle Ages. When there was a shortage of silver for various reasons, medieval craftsmen saved this metal. They made a blank that was covered with silver. There is no way to determine it by other traditional methods, except maybe breaking the coin in half and looking at what's inside.

"We also examined bronze Greek coins from excavations on the Black Sea coast, covered with a very thick layer of corrosion. Using neutron tomography, we were able to see what the original image was like. They studied, for example, finds from an ancient Roman settlement on the territory of Romania, samples of ceramics from the Bronze and Iron Ages, as well as the Middle Ages. And for all these tasks, neural networks were used," Bulat Bakirov emphasizes.

They help to remove noise from the resulting images and to reduce the number of projections that need to be made for high-quality reconstruction of the model. As scientists themselves admit, "completing" a picture using artificial intelligence algorithms allows them to save up to 80% of their time and to explore five times more objects. Our immediate plans are to learn how to improve the quality of reconstructions using neural networks; FLNP scientists currently work on it.

Kseniya MORUNOVA