PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Yuan Gao, Zichen Lu, Xinyi Wu, Wenjun Yu, Shengli Liu, Jianbo Du, Yanliang Jin, Shunqing Zhang, Xiaoli Chu, Shugong Xu

arXiv:2601.00159v1eess.SP

TL;DR

6G CSI acquisition faces overhead and scalability challenges in high-mobility, EL-MIMO, and multi-band systems, while comprehensive coverage of CSI extrapolation research is lacking. This paper surveys model-driven and AI-driven extrapolation across time, frequency, antenna, and multidomain settings, along with datasets, simulators, challenges, and future directions. It positions CSI extrapolation as a way to infer complete CSI from partial CSI while reducing feedback and transmission overhead.

  • Problem

    Conventional channel estimation faces scalability and overhead challenges in high-mobility, EL-MIMO, and higher-frequency 6G systems, while comprehensive CSI extrapolation surveys are lacking.

  • Method

    The paper comprehensively reviews model-driven and AI-driven CSI extrapolation across time, frequency, antenna, and multidomain settings, plus relevant datasets, simulators, challenges, and future directions.

  • Results

    The review synthesizes the principles, strengths, weaknesses, and computational considerations of major CSI extrapolation techniques and identifies open-source resources for training AI models.

  • Takeaways & Limitations

    CSI extrapolation is presented as a key approach for reducing frequent feedback and transmission overhead in future wireless communication systems.

Abstract

from arXiv · show

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.

I. INTRODUCTION

6G CSI acquisition must balance accuracy with low estimation overhead as high mobility, EL-MIMO, higher frequencies, and multi-band operation challenge conventional pilot-based methods. This survey addresses the gap by systematically reviewing CSI extrapolation across domains, model families, datasets, simulators, challenges, and future directions.

  • Motivation: Conventional pilot-based channel estimation becomes poorly scalable in 6G EL-MIMO and higher-frequency contexts, creating prohibitive overhead.CSI extrapolation instead infers complete CSI from a subset of CSI to reduce overhead.
  • CSI extrapolation domains: Time-, frequency-, antenna-, and multi-domain extrapolation target rapidly changing channels, cross-band or reciprocal channels, unmeasured antennas, and jointly unknown CSI.These approaches exploit temporal, cross-frequency, spatial, or multidomain correlations to reduce estimation and feedback burdens.
  • Research gap: Existing surveys lack systematic comparisons of technique strengths, weaknesses, and scenario suitability.They also incompletely cover multidomain extrapolation and generally overlook datasets and channel simulators needed for AI-model training.
  • Scope and contributions: This paper reviews CSI extrapolation fundamentals, representative AI models, and model-driven and AI-driven schemes across time, frequency, antenna, and multidomain settings.The review compares principles, strengths, weaknesses, and computational complexity, with particular attention to multidomain challenges.
  • Scope and contributions: The survey examines open-source wireless datasets and channel simulators and discusses challenges and future directions for AI-driven CSI extrapolation.The proposed directions include dataset construction, performance evaluation metrics, advanced model design, and integration with emerging techniques.

II. FUNDAMENTALS OF AI-DRIVEN CSI EXTRAPOLATION

AI-driven CSI extrapolation builds on a wireless-channel model in which channel coefficients vary across antennas, paths, frequency, and time. In high-mobility settings, historical CSI can be used to extrapolate future CSI, reducing the impact of channel aging and supporting communication design.

  • A. Wireless channel characteristics: The MIMO-OFDM channel coefficient h_i,j(f, t) represents the link between BS antenna i and UE antenna j at frequency f and time t.Its path-level representation uses amplitude, phase, receiving frequency, and transmission delay for each propagation path.
  • A. Wireless channel characteristics: The complete channel matrix H(f, t) organizes the antenna-pair channel coefficients at a given frequency and time.The matrix dimensions are determined by the numbers of BS and UE antennas.
  • 1) Time-domain:: In conventional systems, CSI changes slowly within the coherence time, so pilot-based acquisition can be performed less frequently when mobility is limited.The coherence time decreases as relative mobility increases.
  • 1) Time-domain:: Higher mobility increases Doppler shifts and changes path amplitudes, phases, and potentially the number of paths, producing faster channel variation.These mechanisms make previously acquired CSI become outdated over shorter periods.
  • 1) Time-domain:: Time-domain CSI extrapolation uses historical CSI to predict CSI at future time instances for rapidly changing channels.Short-period extrapolation is driven mainly by Doppler frequency shift, whereas long-period extrapolation also requires location changes; the predictions can support resource allocation, beamforming, and precoding.

2) Frequency-domain:

Frequency-domain CSI extrapolation infers CSI in one frequency band from known CSI in another, supporting FDD and multi-band systems while addressing their distinct propagation characteristics and feedback overhead.

  • FDD separates uplink and downlink bands beyond the coherence bandwidth, so full channel reciprocity does not hold.
  • Partial reciprocity can still permit uplink–downlink inference because propagation-path delays and angles may remain frequency independent while complex gains differ.
  • Multi-band extrapolation targets unobserved bands such as mmWave or THz using observed-band CSI, but their distinct propagation characteristics make the task challenging.
  • Frequency-domain CSI extrapolation uses known CSI from one band to acquire CSI in another, reducing feedback overhead and enhancing spectral efficiency.
  • Antenna-domain CSI extrapolation becomes necessary as EL-MIMO scales, using measured antenna or antenna-pair CSI to infer unmeasured channels and reduce training overhead.

4) Multi-domain:

The reviewed AI architectures support CSI extrapolation through temporal, spatial, and sequence modeling, with recurrent and gated designs targeting channel dynamics and computational trade-offs.

  • 1) MLP:: MLPs use multilayer nonlinear transformations to learn mappings from historical channel inputs to future CSI.Their prediction phase has low computational complexity, but training requires substantial data and depends on input order and architecture.
  • 2) RNN-based models:: RNNs model temporal channel dependencies through recurrent hidden states and can capture nonlinear channel variations more effectively than traditional autoregressive models.However, long sequences can cause gradient vanishing or exploding and require careful hyperparameter tuning.
  • 2) RNN-based models:: LSTMs address long-sequence gradient problems by combining recurrent processing with gating mechanisms and cellular states.The architecture uses forget, input, cell-state, and output-gate operations to regulate information flow.
  • 2) RNN-based models:: GRUs simplify recurrent processing by combining long- and short-term states into one hidden state, reducing parameters and computational complexity relative to LSTMs.Update and reset gates balance retained historical information with short-term channel variations.

3) CNN:

CNNs apply shared convolutional kernels to extract local CSI features efficiently, but their local receptive fields limit modeling of global relationships.

  • 3) CNN:: CNNs generate output feature maps by applying shared convolutional kernels across local regions of an input feature map.The output dimensions depend on input size, kernel size, stride, and padding.
  • 3) CNN:: A convolutional layer has complexity O(Hout×Wout×Cin×Cout×Hk×Wk), determined by output dimensions, channel counts, and kernel dimensions.Cin and Cout denote input and output channels, respectively.
  • 3) CNN:: Weight sharing reduces CNN parameter counts and supports computationally efficient feature representation for high-dimensional data.CNNs exploit translation invariance while extracting local features.
  • 3) CNN:: CNNs inherently struggle to model global relationships because their operations focus on local regions.This limitation is particularly relevant when fine-grained global dependencies matter.

4) GNN:

GNNs perform CSI extrapolation by modeling spatial correlations and local dependencies among antenna or port nodes, but their effectiveness depends on graph quality and data distribution.

  • 4) GNN:: GNNs aggregate neighborhood information to capture topological relationships and local dependencies in CSI data.Their spatial modeling makes them suitable for extrapolating unknown port channels.
  • 4) GNN:: Within AGMAE, the GNN decoder reconstructs unknown-port CSI from an encoder basis vector and known-port CSI while preserving spatial smoothness and local correlations.It learns combination coefficients for reconstructing the entire channel matrix.
  • 4) GNN:: GNN complexity is O(KsNsF + NsF^2), where Ks is the neighbor count, Ns the number of ports, and F the feature dimension.This supports large-scale CSI extrapolation in high-resolution fluidic antenna systems.
  • 4) GNN:: GNN extrapolation accuracy can be constrained when the constructed graph fails to represent genuine channel dependencies.Generalization in dynamic channel environments also requires further optimization.

5) Transformer:

Transformers use self-attention to model long-range CSI relationships and support parallel time-domain extrapolation, but attention cost and data requirements remain substantial.

  • 5) Transformer:: Transformers assign dynamic attention weights to input features and use cascaded encoder-decoder components for sequence processing.Their self-attention mechanism focuses on the most relevant parts of the input sequence.
  • 5) Transformer:: Transformer attention has O(n^2) computational complexity, imposing high resource requirements for long sequences.This cost can limit deployment in resource-constrained CSI extrapolation settings.
  • 5) Transformer:: Transformer-based CSI extrapolation relies on large training datasets and may be less efficient on smaller datasets or simpler tasks.The passage contrasts this limitation with cases where CNNs may perform better.

2) Parametric Channel-Based Approaches:

Parametric channel-based approaches use explicit physical channel models for CSI extrapolation, offering interpretability when assumptions fit but degrading under dynamic or complex conditions. AI-based methods instead learn channel patterns from historical CSI, while model hybrids combine complementary architectures at added design complexity.

  • Parametric Channel-Based Approaches: Parametric approaches estimate physical channel parameters such as amplitude, delay, and Doppler shift under a quasi-static assumption.The assumption treats these parameters as time-varying or slowly varying during the extrapolation period.
  • Parametric Channel-Based Approaches: Explicit physical models provide strong interpretability and work especially well when the channel model matches the scenario.Their performance degrades dramatically when dynamic channels invalidate the quasi-static assumption, and accurate parameter estimation is computationally complex.
  • AI-Based Approaches: AI-based schemes learn complex channel patterns and long-term dependencies from historical CSI, making high-accuracy time-domain extrapolation possible.They are motivated by practical channels whose multipath and Doppler effects evolve too complicatedly for model-driven approaches to match reliably.
  • RNN Series-Based Approaches: RNN models perform sequence-to-sequence CSI extrapolation, while LSTM and GRU variants address limitations in long-term dependency modeling.Reported work finds LSTM models learn long- and short-term dependencies better than baseline RNNs and outperform traditional Kalman and RNN models in cited settings.
  • RNN Series-Based Approaches: RNN-based extrapolation accumulates early prediction errors over later steps and incurs latency for long historical sequences.These limitations are undesirable for high-speed communication despite RNNs supporting variable-length input sequences.
  • Transformer-Based Approaches: Transformers capture complex long-range temporal features and enable parallel computation, but require large datasets and substantial training and inference resources.Their positional encoding is also described as not perfectly suited to time-domain CSI sequences.

A. Model-Driven Approaches

Frequency-domain CSI extrapolation spans parametric and AI-driven methods for FDD and multi-band systems. Parametric methods exploit physical frequency structure, while neural models learn cross-frequency relationships but face scalability, sequence-order, and resource constraints.

  • Parametric Channel-Based Approaches: Parametric methods extrapolate CSI across frequencies using physical properties such as multipath components, path loss, delay spread, and angular scattering.Extrapolation-matrix methods reduce computational complexity relative to cited LS- and MMSE-based methods, while covariance methods can avoid additional downlink training overhead.
  • Parametric Channel-Based Approaches: High-resolution parameter estimation exploits multipath structure to widen frequency-domain extrapolation and support low-feedback-overhead FDD massive MIMO.The cited methods include space-alternating generalized expectation approaches for high-resolution parameter estimation.
  • Parametric Channel-Based Approaches: Parametric frequency-domain methods are interpretable and effective when their models fit, but suffer from model mismatch and high parameter-estimation complexity in wideband multipath channels.These limitations are especially relevant when comprehensive multipath characteristics are difficult to model.
  • MLP-Based Approaches: MLP models learn frequency correlations between subchannels to extrapolate unknown CSI and can reduce training overhead in broadband channel prediction.An MLP-based method is described as fast and efficient while achieving accuracies comparable to benchmark models.
  • MLP-Based Approaches: MLPs have low parameter efficiency, producing large models that limit scalability across multiple bands and require substantial data for generalization.These properties challenge real-time deployment across frequency bands and scenarios.
  • RNN-Based Approaches: RNN-series models exploit temporal information for frequency extrapolation, but sequential prediction accumulates error and frequency correlations are not strictly sequential.One cited method improves RMSE by more than 1 dB over a conventional latest-observation method and maintains accuracy across frequency bands.
  • CNN-Based Approaches: CNNs capture local subcarrier relationships and can learn uplink-to-downlink connections without channel-sparsity assumptions, but struggle with wideband and long-range frequency correlations.They require careful filter design for the structure of wideband frequency-domain data.
  • Transformer-Based Approaches: Transformers capture long-range frequency relationships and are promising for multi-band extrapolation, but their training and inference costs make them unsuitable for resource-constrained scenarios.The cited approaches include end-to-end frameworks that directly map uplink pilots to downlink CSI.

V. ANTENNA-DOMAIN CSI EXTRAPOLATION

Antenna-domain CSI extrapolation predicts CSI at unmeasured antennas using spatial structure. Interpolation and physical models are interpretable and efficient under suitable assumptions, while AI methods improve flexibility but face spatial-modeling and scalability limits.

  • Overview: Antenna-domain extrapolation acquires CSI at antennas of interest from CSI associated with other antennas, reducing estimation demands in MIMO systems.The paper reviews both model-driven and AI-driven schemes for this domain.
  • Model-Driven Approaches: Model-driven methods exploit spatial correlation, adjacent-antenna consistency, and angular-domain sparsity to reconstruct unmeasured antenna CSI.These properties are particularly relevant to massive MIMO and millimeter-wave channels.
  • Interpolation-Based Approaches: Interpolation methods estimate antenna CSI from spatial continuity without requiring explicit physical parameters.Linear interpolation is simple but limited; spline and Kriging methods provide smoother or statistically optimized estimates, while spatial covariance works well in stable environments.
  • Interpolation-Based Approaches: Interpolation methods are computationally efficient and effective for spatially smooth channels, but cannot provide high-accuracy extrapolation when antenna CSI varies non-smoothly.The limitation is highlighted for ultra-massive MIMO systems.
  • Parametric Channel-Based Approaches: Parametric antenna-domain methods extract multipath parameters such as amplitude, delay, and angle of arrival from physical models to predict unmeasured antennas.They use array geometry and spatial correlation, but accurate parameter estimation is computationally complex and real channels may violate model assumptions.
  • AI-Driven Approaches: AI-driven methods learn spatial relationships directly from known CSI, avoiding reliance on explicit physical models for complex antenna configurations.MLPs map partial antenna CSI to full-array CSI through supervised learning, but their parameter inefficiency limits ultra-massive-MIMO scalability.

2) CNN-Based Approaches:

CNN-based antenna-domain methods exploit local spatial structure, whereas GNNs model antenna layouts as graphs and Transformers target unseen configurations. Multi-domain extrapolation can improve accuracy and robustness but increases complexity and synchronization demands.

  • CNN-Based Approaches: CNNs represent CSI as a two-dimensional structure and use convolutional layers to capture local correlations between antennas.They have been used to predict antenna-domain channel statistics and missing antenna CSI.
  • CNN-Based Approaches: CNNs suit grid-like antenna arrays but struggle with irregular layouts, non-stationary spatial characteristics, and long-range antenna dependencies.Graph-based modifications may be needed for fluid or otherwise irregular antenna systems.
  • GNN-Based Approaches: GNNs model antenna panels as graphs, making them suitable for irregular arrays and enabling dependencies across varied antenna layouts.The AGMAE approach supports different CSI input counts and array sizes after training on one fixed-size array.
  • GNN-Based Approaches: The AGMAE method has reported advantages over compressed sensing and linear interpolation in flexibility, computational complexity, and generalization ability.The cited result emphasizes adaptation across input counts and array sizes.
  • GNN- and Transformer-Based Approaches: GNN effectiveness depends on graph design, computational demands, and data availability, while Transformer approaches also require substantial data and resources.Transformer models may fail to capture fine-grained spatial correlations without sufficient training, limiting real-time use.
  • Multi-Domain CSI Extrapolation: Multi-domain CSI extrapolation integrates time, frequency, and antenna dimensions, potentially improving accuracy, robustness, and overhead relative to single-domain extrapolation.Its disadvantages include higher complexity and more difficult synchronization and calibration of multi-domain data.
  • Multi-Domain Model-Driven Approaches: Parametric multi-domain methods estimate multipath parameters before extrapolating across domains, but high-accuracy algorithms are computationally complex and separate domain correlations.Performance degrades when multipath-rich channels cannot be modeled comprehensively.
  • Multi-Domain AI-Driven Approaches: AI-driven multi-domain research includes interleaving-learning MLP mixers that learn antenna and frequency characteristics separately before modeling their cross-domain correlation.The cited design reduces learning burden and improves channel-mapping efficiency, while MLP implementation remains simple but challenging to scale.

1) MLP-Based Approach:

MLP- and CNN-based approaches exploit learned channel correlations for multi-domain CSI extrapolation, but their effectiveness depends on scenario assumptions, locality, and computational cost. The section also connects AI performance to the availability and suitability of open-source datasets.

  • MLP-Based Approach:: MLP-based methods combine data-driven networks with prior assumptions about physical channel characteristics, so performance depends on how well those assumptions match the target scenario.
  • CNN-Based Approach:: 3D CNNs capture temporal, spatial, and frequency correlations for multi-domain CSI extrapolation, while CNN locality limits modeling of broader channel dependencies.
  • CNN-Based Approach:: CNN-based approaches can be computation-intensive, creating deployment challenges for joint high-speed, EL-MIMO, and wideband systems.
  • Transformer-Based Approach:: Transformers are motivated by the need for better generalization in complex communication scenarios, including models that jointly represent temporal, spatial, and frequency features.
  • Generative and Hybrid AI:: Generative AI may improve generalization across diverse scenarios, but large generative architectures require substantial computational resources and training data.
  • Generative and Hybrid AI:: Hybrid schemes combine complementary models; one GAN-LSTM framework generates synthetic channel data to enrich training before sequence prediction.
  • Datasets:: Open-source datasets span measured and ray-tracing-generated channels, with different strengths across temporal, frequency, antenna, mobility, and scenario coverage.
  • Datasets:: Industrial Radio and Wireless Intelligence provide 1 kHz measurements for time-domain studies, whereas DICHASUS offers 32 antennas and 1024 subcarriers for spatial and frequency-domain research.

B. Channel simulators

The surveyed channel simulators cover diverse 6G frequencies, scenarios, antenna configurations, and propagation domains. Their capabilities range from calibrated statistical and link-level models to high-fidelity, near-field, RIS, and differentiable ray-tracing tools.

  • Channel simulators: SEU-PML-6GPCS is a calibrated unified geometric-stochastic simulator supporting multi-user, cross-band, and diverse 6G scenarios including satellite, UAV, and terrestrial communication.
  • Channel simulators: It can model arbitrary transceiver and scatterer motion while saving channel characteristics across time, frequency, spatial, delay, Doppler, and angle domains.
  • Channel simulators: BUPTCMCCCMG-IMT2030 generates channel parameters and coefficients across scenarios, bands, and antenna arrays through configuration, simulation, and analysis modules.
  • Channel simulators: NYUSIM is an open-source statistical model supporting 0.5–150 GHz, multiple 3GPP scenarios, polarization configurations, dynamic blockage, and ns-3 integration.
  • Channel simulators: Sionna RT provides GPU-accelerated differentiable ray tracing, enabling gradient-based optimization with respect to environmental, antenna, and material parameters.
  • Channel simulators: QuaDRiGa is a measurement-calibrated system-level simulator supporting standardized models and applications including indoor, urban, V2X, and satellite scenarios.
  • Channel simulators: WiThRay targets physically accurate RIS and metasurface environments across sub-6 GHz, mmWave, and sub-THz bands.
  • Channel simulators: NirvaWave uses scalar diffraction and Fourier-optics principles for near-field propagation, including RIS phase shifts and rough-scattering effects.

1) Time-domain:

Time-domain CSI extrapolation faces growing difficulty as mobility, prediction horizons, and environmental dynamics increase. The section identifies dataset and simulator limitations and proposes hybrid data and simulation strategies.

  • 1) Time-domain:: Short-term time-domain extrapolation historically relied on near-stationarity and Doppler-dominated variation, but 5G-Advanced and 6G target speeds up to 1000 km/h.
  • 1) Time-domain:: Long-term extrapolation is constrained because stochastic models capture Doppler fading but struggle with location-dependent propagation, while ray tracing does not easily model continuous trajectories.
  • 1) Time-domain:: Most simulators omit moving objects, temporary blockages, foliage motion, phase noise, and oscillator drift, making reported extrapolation accuracy potentially optimistic.
  • 1) Time-domain:: Cross-band extrapolation is harder than same-band FDD extrapolation because path loss, multipath richness, and scattering mechanisms vary substantially with frequency.
  • 1) Time-domain:: Near-field antenna-domain extrapolation is complicated by spherical wavefronts, changing path visibility, and spatial non-stationarity across large apertures.
  • 1) Time-domain:: Realistic near-field datasets are costly and difficult to collect, and existing methods struggle with discontinuous path visibility rather than smooth spatial correlation.
  • 1) Time-domain:: Multi-domain errors interact and amplify, so models trained or evaluated only on simulated data can show good single-domain results but severe joint prediction errors.
  • Future research: Hybrid datasets combining high-fidelity simulations with targeted measurements are proposed to provide diverse, physically consistent multi-domain samples.

2) Comprehensive Performance Evaluation: •

The section frames generalization, data requirements, computational efficiency, architecture design, multimodal inputs, and emerging 6G technologies as central evaluation and research concerns for CSI extrapolation.

  • Comprehensive Performance Evaluation:: Generalization is a critical deployment metric because differing measurement setups and scattering environments can create severe distribution shifts across datasets.
  • Comprehensive Performance Evaluation:: The required dataset volume remains insufficiently quantified across CSI extrapolation domains, with needs varying by model, scenario complexity, dynamics, and propagation condition.
  • Comprehensive Performance Evaluation:: High-complexity Transformers and deep CNNs may improve accuracy but can be impractical for latency-sensitive, resource-constrained applications such as V2X.
  • Comprehensive Performance Evaluation:: Sparse-attention Transformers reduce self-attention complexity toward near-linear or logarithmic levels, supporting lower-complexity real-time CSI extrapolation.
  • Comprehensive Performance Evaluation:: Model-driven methods improve resilience by combining physical channel knowledge with data-driven techniques, while Transformer, sparse-attention, and mixture-of-experts designs are proposed for future models.
  • Comprehensive Performance Evaluation:: Multimodal learning is proposed because geometry, mobility, and sudden blockages cannot be fully captured by CSI alone.
  • Comprehensive Performance Evaluation:: XL-MIMO, RIS, and FAS introduce spatial non-stationarity, controllable time variation, and higher dimensionality, increasing real-time extrapolation demands on constrained devices.
  • Comprehensive Performance Evaluation:: NOMA and RSMA impose higher CSI accuracy and timeliness requirements because they deliberately introduce controlled inter-user interference.
Loading 2601.00159v1…