This paper presents a small, efficient artificial intelligence (AI) method to estimate radio channels from handheld users to low‑Earth‑orbit (LEO) satellites. The authors aim to run channel estimation on the satellite itself, where power and computation are limited. They show a convolutional neural design that reduces work and memory use while improving accuracy, measured by mean squared error (MSE).

The researchers built on a low‑complexity neural receiver called MDX and added a new building block named MDELAN (Multi‑Dilated Efficient Layer Aggregation Networks). The overall receiver first forms two simple estimates from pilot tones and from tentative decoded data. Those estimates, together with a positional encoding that tells the network each subcarrier and symbol location, are fed to the neural network. The MDELAN blocks use dilated convolutions to see wider patterns in time and frequency while keeping the number of parameters small by using depthwise separable convolutions.

They tested the models in simulations of 6G uplink physical channels using the QuaDRiGaNT channel generator. Simulation details include an S‑band carrier at 2 GHz, an OFDM uplink frame with 14 symbols per transmission time interval (about 500 µs), 30 kHz subcarrier spacing, QPSK modulation, and pilot symbols placed in OFDM symbols 2 and 11. The scenarios included rural, suburban and urban NTN (non‑terrestrial network) environments and user speeds up to 100 km/h. Performance was compared to standard baselines: least squares (LS) with linear interpolation and linear minimum mean squared error (LMMSE) interpolation.

Why this matters: LEO satellites face tight limits on power, memory and onboard processors. Conventional high‑accuracy methods like LMMSE need heavy computation and prior channel statistics, which can be hard to get or too costly for a satellite. The proposed design is targeted for real‑time onboard inference. The authors report that, by exploiting channel structure, their design improves parameter efficiency by about 27% compared with state‑of‑the‑art AI models and requires roughly 29× fewer floating‑point operations than conventional methods while achieving superior MSE in their tests. They also provide a complexity analysis in terms of multiply‑and‑accumulate operations, floating point operations and learnable parameters.