Large-Scale AI CSI Compression for 6G FR3

In the 6G FR3 band (7.125–24.25 GHz), X-MIMO with hundreds of antenna ports (such as 256 ports) is required, which makes CSI feedback emerge as a key bottleneck. Existing codebook methods are difficult to scale to wideband channels of this size, and training an AI autoencoder from scratch becomes unstable under extreme compression conditions.

To address this, Samsung Research proposes a two-stage training method.

• Stage 1 – HTL-PC (Hierarchical Transfer Learning + Progressive Compression): Gradually increases the compression ratio while transferring knowledge from each stage to the next, enabling stable training of a deep decoder.
• Stage 2 – DGET (Decoder-Guided Encoder Training): Freezes the deep decoder trained via HTL-PC, then separately trains only a lightweight 1-layer encoder so that it produces latent representations the decoder already understands.

As a result, the DGET-trained 1-layer encoder achieves an approximately 18% improvement in SGCS over the eType II codebook, with only a 3.5–7.2% accuracy loss compared to a 7-layer encoder, while cutting UE-side computation by 82–84%.

Angular domain analysis confirmed that even the 1-layer encoder sufficiently captures the energy of major beam directions. The method was also validated by applying a model trained solely on synthetic data to real-world OTA testing without additional fine-tuning, and this research is directly tied to 3GPP Release 20 standardization of AI-based CSI feedback.