5G-NR TDL Channel Dataset for Deep Learning

he dataset was generated using the MATLAB Data Synthesis for Channel Estimation in fifth-generation (5G) code (DL_Matlab_5G-NR) . A summary of the parameters used to generate the dataset, such as the type of physical channel (e.g., Physical Downlink Shared Channel (PDSCH)), central frequency, Subcarrier Spacing (SCS), Cyclic Prefix (CP) type, number of RBs, code rate, and modulation, among others, is provided in this paper .

Compared to the original code, some minor modifications were made. Specifically, the SNR values ranged between 0 and 20
dB, instead of 0 to 10 dB, and we generated 1024 samples for each SNR value, resulting in a total of 11,264 samples (11
SNR values × 1024 samples per value), with a split of 70% for training, 15% for validation, and 15% for testing. The true
labels are generated by assuming full CSI, while the model inputs are of two types: LS estimates on the pilot symbols, for
HELENA and LSiDNN, and LI after the initial LS estimates, as utilized in ChannelNet, EDSR, AttRNet, and CEViT.