ChanEst Dataset for Deep Learning-Based 6G Channel Estimation

ChanEst dataset was developed to address the core research challenge that accurate channel estimation is increasingly difficult due to highly diverse, dynamic, and extreme propagation environments in 6G. Although deep learning (DL) techniques have shown strong potential as alternatives to conventional estimators, their progress is limited due to the lack of reproducible and reconfigurable datasets that adhere to 3GPP compliance and realistic receiver preprocessing. ChanEst directly responds to this gap by providing a fully controlled, standards aligned dataset generation framework designed for DL based channel estimation.

ChanEst is generated using the 6G Exploration Library in MATLAB and strictly follows 3GPP physical layer specifications. Each sample is constructed on an OFDM grid (612 subcarriers × 14 symbols) populated with DM RS Type 2 pilots (3GPP TS 38.211). The transmitted signals propagate through standardized 3GPP TDL channel models (TDL A to TDL E), while channel parameters, SNR (−10 to 30 dB), delay spread (10–2000 ns), and Doppler shift (5–5000 Hz), are stratified to ensure wide and balanced scenario diversity representative of 6G FR3 operation at 7 GHz. This design prevents overrepresentation of mild or unrealistic cases and ensures coverage of high mobility and high dispersion conditions.

Each dataset entry consists of two tensors:
• X_input: LS channel estimates at pilot positions followed by 2 D linear interpolation, representing the noisy and imperfect receiver perspective.
• Y_label: perfect OFDM grid channel responses, serving as the supervised learning target.
Both tensors are stored in real valued format, with channels represented as [K × L × C × N], where C = 2 × NTx × NRx to pack real and imaginary components. The framework supports SISO and MIMO, configurable numerologies, and user defined parameter ranges. Dataset size is scalable based on computational capacity, making ChanEst suitable for lightweight prototyping as well as large scale model training.

Extensive validation confirms that ChanEst exhibits strong physical consistency. Correlation between X_input and Y_label is high under mild channel conditions and degrades appropriately under severe scenarios, reflecting real world estimation difficulty. NMSE trends align with expected impacts of noise, delay spread, and mobility, demonstrating that the dataset avoids degenerate or trivial cases. Scenario distributions remain balanced across propagation regimes, ensuring fairness and robustness in DL model benchmarking.

ChanEst is delivered with detailed metadata, including SNR, Doppler, delay spread, and TDL profile, enabling stratified evaluation, targeted stress testing, and reproducible comparisons against classical estimators. Its flexibility allows researchers to reconfigure antenna dimensions, numerology, and parameter ranges to suit diverse DL-based research tasks in 6G and beyond. The dataset is available in MATLAB (.mat) and HDF5 (.h5) formats.