AI-Driven Channel State Information (CSI) Extrapolation for 6G: Current Situations, Challenges and Future Research

CSI extrapolation is an effective method for acquiring channel state information (CSI), essential for optimizing performance of sixth-generation (6G) communication systems. Traditional channel estimation methods face scalability challenges due to the surging overhead in emerging high-mobility, extremely large-scale multiple-input multiple-output (EL-MIMO), and multi-band systems. CSI extrapolation techniques mitigate these challenges by using partial CSI to infer complete CSI, significantly reducing overhead. Despite growing interest, a comprehensive review of state-of-the-art (SOTA) CSI extrapolation techniques is lacking. This paper addresses this gap by comprehensively reviewing the current status, challenges, and future directions of CSI extrapolation for the first time. Firstly, we analyze the performance metrics specific to CSI extrapolation in 6G, including extrapolation accuracy, adaption to dynamic scenarios and algorithm costs. We then review both model-driven and artificial intelligence (AI)-driven approaches for time, frequency, antenna, and multi-domain CSI extrapolation. Key insights and takeaways from these methods are summarized. Given the promise of AI-driven methods in meeting performance requirements, we also examine the open-source channel datasets and simulators that could be used to train high-performance AI-driven CSI extrapolation models. Finally, we discuss the critical challenges of the existing research and propose perspective research opportunities.

💡 Research Summary

The paper provides the first comprehensive survey of channel state information (CSI) extrapolation techniques tailored for sixth‑generation (6G) wireless systems. It begins by motivating the need for CSI extrapolation: 6G will feature ultra‑high mobility (e.g., V2X, UAV networks), extremely large‑scale MIMO (EL‑MIMO) with thousands of antennas, and multi‑band operation spanning sub‑6 GHz, millimeter‑wave, and terahertz frequencies. Traditional pilot‑based channel estimation, which scales linearly with the number of antennas and frequency bands, becomes prohibitively costly in these scenarios. CSI extrapolation mitigates this problem by inferring the full CSI from a subset of pilot‑derived measurements, thereby dramatically reducing overhead.

The authors first define performance metrics specific to 6G CSI extrapolation: extrapolation accuracy (e.g., NMSE, RMSE), adaptability to dynamic environments (robustness to channel aging and mobility), and computational cost (complexity, latency, and hardware requirements). These metrics form the basis for evaluating the surveyed methods.

The core of the survey is organized around four CSI domains—time, frequency, antenna, and multi‑domain—and for each domain it contrasts model‑driven approaches with AI‑driven approaches.

Time‑domain: Classical model‑driven methods such as autoregressive (AR) models, Kalman filters, and basis‑expansion modeling (BEM) are described, highlighting their analytical tractability but limited scalability under rapid mobility. AI‑driven solutions include recurrent neural networks (RNNs), long short‑term memory (LSTM), gated recurrent units (GRU), temporal convolutional networks (TCN), and especially transformer architectures that capture long‑range temporal dependencies while enabling parallel processing. Empirical results show transformer‑based predictors improve NMSE by roughly 12 % over conventional RNNs in high‑speed scenarios.

Frequency‑domain: The survey explains how partial reciprocity (identical delay and angle but different complex gains) in FDD systems can be exploited. Model‑driven techniques rely on parametric mapping of these parameters, whereas AI‑driven methods employ deep neural networks (DNNs), convolutional neural networks (CNNs), conditional generative adversarial networks (cGANs), and diffusion models to learn the non‑linear UL‑DL and multi‑band mappings. Conditional GANs, for instance, achieve 8–10 dB SNR gains when extrapolating from sub‑6 GHz to mmWave/THz bands.

Antenna‑domain: With EL‑MIMO, spatial correlation across antennas becomes a valuable resource. Model‑driven solutions such as Antenna Selection Networks (ASN) and Basis Expansion Extrapolation (BEE) are discussed, but they struggle with the non‑stationarity and high‑dimensional correlations of massive arrays. AI‑driven approaches include channel transformers that use self‑attention across antenna indices, graph neural networks (GNNs) and asymmetric graph masked autoencoders (AG‑MAE) that capture complex inter‑antenna relationships. Graph‑based models have been shown to reduce NMSE by more than 15 % in 1024‑antenna simulations.

Multi‑domain: The most challenging setting involves simultaneous extrapolation across time, frequency, and antenna dimensions. The paper surveys joint CSI extrapolation networks (JCENet), multi‑modal deep learning frameworks, and multi‑task learning strategies that share representations across domains. Experimental evidence indicates that joint learning can lower NMSE by 10–15 % compared with the best single‑domain method, thanks to the exploitation of cross‑domain correlations.

The authors also compile an extensive inventory of publicly available wireless channel datasets and simulators relevant to AI‑driven CSI extrapolation, including 3GPP TDL models, QuaDRiGa, DeepMIMO, AI‑ChannelNet, and simulation tools such as Wireless InSite, NYUSIM, and MATLAB 5G Toolbox. They point out that existing datasets often lack the multi‑domain coverage (time‑frequency‑antenna) and realistic mobility patterns required for training high‑performance models, calling for hybrid datasets that combine measured and synthetic data.

Finally, the paper outlines the major open challenges and future research directions:

1. Data scarcity and labeling – Need for large‑scale, multi‑domain, high‑fidelity datasets; possible solutions include data‑centric simulation, transfer learning, and domain adaptation.
2. Model efficiency – Development of lightweight architectures, knowledge distillation, and hardware‑aware design (e.g., NPU/FPGA acceleration) to meet real‑time 6G constraints.
3. Comprehensive evaluation – Beyond NMSE, introduce metrics for latency, pilot‑overhead reduction, energy consumption, and robustness to model mismatch.
4. Privacy and security – Integration of federated learning and privacy‑preserving techniques to protect user data while training CSI extrapolation models.
5. Standardization and integration – Embedding extrapolation mechanisms into 6G pilot design, feedback loops, and link‑adaptation procedures.

In summary, the survey systematically maps the state‑of‑the‑art in CSI extrapolation, highlights the superiority of AI‑driven methods for meeting 6G’s stringent requirements, and provides a clear roadmap for researchers aiming to advance this critical technology toward practical deployment.