Scientists used neural network to reconstruct 3D images of fibrous materials

Scientists used neural network to reconstruct 3D images of fibrous materials
Scientists used neural network to reconstruct 3D images of fibrous materials
Anonim

Skoltech researchers and their colleagues at the Catholic University of Leuven (Belgium) managed to reconstruct 3D images of fibrous materials obtained using microcomputer tomography. To solve this complex and time-consuming task for a person, scientists used machine learning methods.

Image

The results obtained, published in the journal Computational Materials Science, are important for further in-depth analysis of material properties. Microcomputed tomography is an indispensable tool for studying the three-dimensional microstructure of fiber-reinforced composites and other complex materials. However, the use of this method is associated with a number of additional difficulties, such as very small sample sizes, the presence of artifacts and shaded areas in the images, as well as low quality or complete absence of individual image fragments.

To solve this difficult task, scientists decided to use the methods that restorers use when restoring works of art - in particular, the method of reconstructing defects, which is already widely used in digital image processing.

“The main advantage of AI-based image reconstruction is speed. With a trained model, this method can process up to hundreds of images per second. Such a speed is simply beyond the power of man. In addition, computers are much better at processing three-dimensional images, since the machine is able to see them 'through and through' and from all sides, instantly carrying out reconstruction over the entire volume, and not just over the surface,”says the first author of the article, a graduate student of Skoltech and Leuven Catholic University Radmir Karamov.

Karamov participates in research led by the director of the Skoltech Center for Design, Manufacturing Technologies and Materials (CDMM) Professor Iskander Akhatov and Professor Stepan Lomov of the Catholic University of Leuven. The team proposed to use generative adversarial networks (GANs) with 3D encoders and decoders for the tasks of reconstructing three-dimensional images of microcomputer tomography.

As the authors explain, reinforcing inclusions in composite materials, such as fibers, can have an arbitrary orientation in three dimensions, therefore it is necessary to deal with 3D images describing this complex internal microstructure. Since it was not possible to achieve the required accuracy with the help of more familiar convolutional neural networks, the scientists decided to use generative adversarial networks.

“When reconstructing images using GANs, it is necessary to train for this purpose not one, but two competing neural networks: a generative network that forms 'artificial' images that look like genuine ones; and a discriminative network whose task is to distinguish “real” images from “artificial” ones.

As GAN creator Ian Goodfellow said, this is reminiscent of the rivalry between counterfeiters and cops.The former strive to print as many counterfeit bills as possible, which differ little in appearance from the real ones, while the latter check each bill for authenticity,”explains Karamov. Scientists tested three versions of the GAN architecture, choosing for this purpose microcomputed tomography images of the most complex sample - a composite reinforced with short glass fibers, which does not have any repetitions in its structure.

As a result, out of three options, the researchers chose a network architecture that most successfully combined high quality reconstruction, performance, and moderate use of GPU memory. “Our proposed algorithm allows us to eliminate all defects in images and, therefore, more accurately model the properties of materials and predict the quality of the final material, provided that all internal pores and voids in its structure are eliminated during the production process,” emphasizes Karamov.

Reconstruction of the microstructure of materials is the first step in developing a fully automatic generative algorithm that will allow the creation of innovative materials with properties that meet the requirements of specific applications, he adds.

Popular by topic