Russian scientists have developed a neural network designed for calculating holograms.

Typically, computationally intensive iterative methods are employed to calculate (synthesize) diffractive and holographic optical elements that can optically form a 3D scene. Researchers at MEPhI have proposed a neural network method called 3D-CGH-Net, which allows for rapid calculation of optical elements. It has been experimentally demonstrated that the quality of optically reconstructed 3D scenes from these elements can surpass that of elements synthesized using traditional, more resource-intensive algorithms.

“The speed of calculating high-quality holograms using ‘conventional’ methods is low,” explains laboratory employee Dmitry Rymov. “We have developed a method that utilizes a neural network with an original architecture and a branched structure to account for a large set of cross-sections of the three-dimensional scene in the calculated hologram. The network is trained on datasets ranging from tens of thousands to hundreds of thousands of examples. The method has been successfully applied in experiments involving the optical formation of three-dimensional scenes, as well as in the implementation of holograms using state-of-the-art high-resolution, high-speed spatial-temporal light modulators.”

According to the lab head, Professor Rostislav Starikov, computer-generated holography involves calculating holograms that can then be realized in various ways, such as through specific materials, printing, or using spatial-temporal light modulators (essentially micro-displays). The application of computer-synthesized holograms allows for the precise and rapid formation of specified light distributions (including those that do not exist in nature).

“The use of computer-generated holograms is promising for creating three-dimensional visualization tools, for laser manipulation of microparticles, in photostimulation of biological neurons, for 3D printing, in transforming and focusing light beams, in constructing holographic memory systems, and much more,” explained Rostislav Starikov.

The technique of computer-generated holography has been evolving since the late 1960s and is now quite sophisticated. To calculate computer-generated holograms, it is necessary to solve the inverse problem (calculate the shape of the diffractive element based on the required amplitude and phase distribution of light it generates). There are several “classical” methods for this, which are generally iterative and computationally intensive; the calculation of a hologram can take hours, which is often unacceptable in modern practice. “The application of neural networks allows for the calculation, or more accurately, the generation of a hologram if the network has been successfully trained beforehand. Training requires time and large datasets, but a trained network generates a hologram very quickly,” asserts the lab head.

“The latest intelligent methods already allow us to significantly expand the boundaries of neural network applications, addressing not only amateur tasks but also complex scientific problems,” notes Pavel Cheremkhin, an associate professor at the Department of Laser Physics at MEPhI. “In just a couple of years, we have achieved such high-quality calculations of holograms using neural network methods that they exceed some capabilities of standard methods that have been developed over decades.

Our laboratory has developed a method that synthesizes megapixel holograms of complex three-dimensional scenes in just fractions of a second. Additionally, a high level of optical reproduction quality for these 3D scenes from such holograms has been achieved. The use of modern spatial-temporal light modulators allows for the formation of thousands of holograms per second, enabling the alteration or transformation of three-dimensional light distributions thousands of times per second.”

The results of the research are presented in the high-ranking journal Optics and Lasers in Engineering.