CSI Prediction Using Diffusion Models

Mehdi Sattari, Javad Aliakbari, Alexandre Graell i Amat, Tommy Svensson

TL;DR

This paper introduces a novel probabilistic framework for Channel State Information (CSI) prediction using diffusion models. By shifting from deterministic mappings (like RNNs or Transformers) to generative modeling, the authors successfully capture the stochastic nature of wireless channels, outperforming state-of-the-art baselines by up to 8 dB in prediction accuracy while maintaining robustness across different mobility and frequency scales.

What Problem Does This Paper Address?

In modern MIMO (Multiple-Input Multiple-Output) systems, the base station needs accurate CSI to direct signals efficiently toward users. However, wireless channels change rapidly due to user mobility and environmental shifts—a phenomenon known as channel aging. By the time CSI is estimated and used, it may already be stale.

While Deep Learning (DL) models like LSTMs and Transformers have been used to predict future CSI from past observations, they typically learn a deterministic mapping. Wireless channels are inherently stochastic; a single past observation could lead to multiple plausible future states. Deterministic models struggle with this "multimodality" and often overfit to specific training distributions, failing when the environment changes (e.g., a change in carrier frequency or user speed).

Methods

The authors propose a flexible, two-part architecture for CSI prediction:

  1. Temporal Encoder: A module (like a ConvLSTM or LinFormer) that extracts the underlying dynamics and "latent features" from historical CSI data.
  2. Diffusion-Based Generator: A generative backbone (such as a U-Net or Diffusion Transformer) that takes those latent features and "denoises" a Gaussian noise signal into a precise prediction of future CSI.

The researchers explored three primary configurations:

  • AR (Autoregressive): Predicts the next CSI frame and feeds it back into the model to predict the one after that. This is flexible for any prediction length.
  • Seq2seq (Sequence-to-Sequence): Predicts an entire block of future frames in one parallel pass. This is faster and avoids "error accumulation" but is less flexible regarding input/output lengths.
  • Unified Seq2seq: A simplified version where a 3D U-Net learns spatial and temporal features simultaneously without a separate encoder.

To make these models practical for low-latency wireless systems, they utilized DDIM (Denoising Diffusion Implicit Models) scheduling, which allows the model to generate high-quality samples in as few as 3 to 5 steps, compared to the 1000 steps often required by standard diffusion models.

Datasets

The models were evaluated using 3GPP-compliant Clustered Delay Line (CDL) channel models (scenarios CDL-A through CDL-E).

  • Training Data: 100,000 independent CSI samples generated at 28 GHz (mmWave).
  • Parameters: 16x1 MIMO system, 300 subcarriers, and user velocities ranging from 30 to 120 km/h.
  • Generalization Test: A separate test set was generated at 3 GHz to see how well the models handle "domain shift" (moving from mmWave to sub-6 GHz frequencies).

Key Insights and Ideas

  • Probabilistic is Better: By modeling the full probability distribution $p(H_{future} | H_{past})$, diffusion models act as a data-driven approximation of the optimal Minimum Mean Squared Error (MMSE) predictor.
  • The "Aha" Moment on Generalization: Deterministic models like LinFormer tend to "memorize" the specific statistics of the 28 GHz training data. When tested at 3 GHz, their performance collapses. Diffusion models, however, learn the structural "physics" of how channels evolve, making them significantly more robust to changes in carrier frequency.
  • Efficiency Trade-offs: While the AR approach (DiU) provides the best accuracy for short-term predictions, the Seq2seq approach (DiU-seq2seq) is much more computationally efficient for long-term horizons because it predicts everything at once.

Key Results

  • Accuracy: At an SNR of 20 dB, the diffusion-based AR model (DiU) achieved an NMSE (Normalized Mean Squared Error) improvement of 5 dB over ConvLSTM and 8 dB over LinFormer.
  • Mobility: The models maintained a lead across all tested speeds (30, 60, and 120 km/h), though the gap narrows at very high speeds where temporal correlation is naturally weaker.
  • Sampling Speed: Using DDIM, the authors showed that prediction accuracy saturates quickly; there is almost no performance difference between using 5 sampling steps and 100 steps, making real-time implementation much more feasible.
  • Qualitative Success: Visualizations showed that while baselines like LinFormer struggle to recover the spatial structure of the channel, diffusion models produce "clean" reconstructions that closely match the ground truth.

Limitations and Future Work

  • Computational Complexity: Despite the reduction in sampling steps, diffusion models are still heavier than simple GRUs. The 3D U-Net variant, for instance, requires over 48 billion FLOPs per pass, which might be too taxing for current mobile hardware.
  • Fixed Contexts: The Seq2seq models are currently tied to fixed input/output lengths. Future work could involve making these more adaptive.
  • End-to-End Integration: The paper focuses on prediction accuracy (NMSE). The next logical step is to evaluate how these predictions directly impact system-level metrics like bit error rate (BER) or overall spectral efficiency in a live beamforming loop.

Summary updated

Abstract

Acquiring accurate channel state information (CSI) is critical for reliable and efficient wireless communication, but challenges such as high pilot overhead and channel aging hinder timely and accurate CSI acquisition. CSI prediction, which forecasts future CSI from historical observations, offers a promising solution. Recent deep learning approaches, including recurrent neural networks and Transformers, have achieved notable success but typically learn deterministic mappings, limiting their ability to capture the stochastic and multimodal nature of wireless channels. In this paper, we introduce a novel probabilistic framework for CSI prediction based on diffusion models, offering a flexible design that supports integration of diverse prediction schemes. We decompose the CSI prediction task into two components: a temporal encoder, which extracts channel dynamics, and a diffusion-based generator, which produces future CSI samples. We investigate two inference schemes-autoregressive and sequence-to-sequence- and explore multiple diffusion backbones, including U-Net and Transformer-based architectures. Furthermore, we examine a diffusion-based approach without an explicit temporal encoder and utilize the DDIM scheduling to reduce model complexity. Extensive simulations demonstrate that our diffusion-based models significantly outperform state-of-the-art baselines.
BibTeX
@article{sattari2026csi,
  author = {Mehdi Sattari and Javad Aliakbari and Alexandre Graell i Amat and Tommy Svensson},
  title = {{CSI Prediction Using Diffusion Models}},
  journal = {IEEE Trans. Wireless Communications},
  year = {2026},
  doi = {10.1109/twc.2026.3719116},
}

Loading related content...

Figure