Enabling Data-Driven OTFS Modulation for 6G Hyper Reliable High-Mobility Communications

Orthogonal time frequency space (OTFS) modulation has emerged as a promising solution for hyper-reliable high-mobility wireless communications over channels with severe Doppler effect. By mapping symbols into the delay-Doppler (DD) domain, the channel effect of OTFS becomes a two-dimensional circular convolution between symbols and channel responses. However, the physical layer algorithms of OTFS are challenging due to the complexity of two-dimensional signal processing. To this end, this article explores the potential of data-driven approaches for OTFS systems since deep learning (DL) is powerful in extracting both linear and nonlinear complex relationships from high-dimensional data. The DD domain signals are easily transferred to grid, graph, or sequential data that are compatible with various DL models. We investigate how these data-driven approaches can be applied to OTFS communication systems, including transmitter and receiver. Furthermore, we propose shift padding (SP) for the convolutional neural network (CNN) to address its inability to handle edge symbols due to the limited reception field. We validate the effectiveness of SP-enhanced CNN for OTFS channel equalization on vehicular channels, and the results show that, for low signal-to-noise ratios, the data-driven approach can achieve higher reliability compared with the conventional algorithms.