Generation of Orthogonal THz Pulses for Wireless Communications Based on Diffractive Autoencoder Neural Networks

Abstract

Pulse-based systems provide a promising alternative to terahertz (THz) communications, especially for joint communication and sensing applications. For such a system, THz Gaussian pulses are the fundamental and commonly used waveforms, but are susceptible to distortion due to their large bandwidth occupation. A critical yet unresolved issue is generating THz pulses with tunable center frequencies and bandwidths. In this paper, a THz-pulse generator is designed based on diffractive surfaces cascaded in multiple layers. Given the THz Gaussian pulse as an input, each surface has the ability to change its amplitude and phase and diffracts it to the next surface, so as to generate pulses with expected frequencies and bandwidths. To determine parameters for millions of elements on all surfaces and to handle the case of generating multiple THz pulses corresponding to the same THz Gaussian input signal, a diffractive autoencoder neural network (DANN) is developed. Subsequently, using the generated pulses for data transmission under THz channel, the symbol error rate (SER) performance is analyzed. Extensive simulations are conducted to validate and evaluate the DANN-based THz pulse generator. Additionally, using just 5 diffractive surfaces, the generator can support at least 10 pairs of orthogonal THz pulses with a correlation error ratio of less than $10^{-1}$ .