Terahertz (THz) optics often encounters the problem of small f-number values (elements have relatively small diameters comparing to focal lengths). The need to redirect the THz beam out of the optical axis or form particular intensity distributions resulted in the application of iterative holographic methods to design THz diffractive elements. Elements working on-axis do not encounter significant improvement while using iterative holographic methods; however, for more complicated distributions the difference becomes meaningful. Here, we propose a totally different approach to design THz holograms, utilizing a neural network based algorithm, suitable also for complicated distributions.

Fast developing terahertz (THz) technology finds applications in many fields of science and industry such as optics, medicine, security, or non-destructive testing [1‒3]. Each new application requires specific optical elements that will allow for full use of the potential of THz radiation under given conditions. Thus, new solutions allowing to meet market needs are still requested. Nowadays, thin diffractive optical elements (DOEs) are used instead of thick refractive lenses. The continuous phase changes, kinoform coding method, allow to redirect the whole electromagnetic radiation into a single diffraction order, which significantly increases the efficiency of the THz diffractive structures [4‒5]. More advanced DOEs can be designed as synthetic holograms. Synthetic holography is based on calculating complex transmittance that, when illuminated, redirects the radiation to form the image of an object Therefore, designed wavefront shapes can be more advanced and complex compared to classic holography [5] or designing DOEs using analytical equations. Regrading applicational requirements, synthetic holography mostly utilizes different kinds of iterative methods like the Gerchberg-Saxton, also called the Iterative Fourier transform Algorithm (IFTA), or the ping-pong algorithms [6‒7]. Such algorithms perform a number of iterations propagating the wavefront between the hologram and the image planes. In each plane a particular amplitude distribution is forced, while the phase is being optimized. Thus, the generated phase distribution in the hologram plane results in forming the