acceptodds
Withdrawn from ICLR 2027

Achieving Robust Channel Estimation Neural Networks on Doppler effects by Designed Training Data

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Channel estimation is crucial in wireless communication systems. Many existing works train and test their models using the same or similar channel. However, practical channels are infinitely diverse and time-variant while data-driven algorithms often degrade on new and previously unseen channels. This prohibits the practical implementation of neural networks for channel estimation, particularly in high speed scenarios characterized by rapidly channel profile changes and very high Doppler shifts. Online training can fine-tune models to adapt to new channels, but requires high computation resource and incurs significant latency for physical-layer devices. Moreover, online training cannot exhaustively cover all possible channels because there are infinitely many possible channels. Therefore, we investigate the ability of neural networks to generalize to wireless channels. This paper proposes design criteria to generate synthetic training data, which guaranty that the resulting networks perform robustly across wireless channels, especially for very high Doppler shifts and different Doppler spectra. It ensures that offline-trained neural networks require no prior information or online training to precisely estimate the new channels. To demonstrate general applicability, we use neural networks with different levels of complexity to show that the generalization achieved appears to be independent of architecture. From simulation results, neural networks achieve a certain performance across wireless channels with different unseen power delay profiles (PDP) and Doppler shifts.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Channel estimation, generalization, neural network, orthogonal frequency division multiplexing (OFDM), wireless communciations

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_BmBMOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes criteria for generating synthetic training data to improve the robustness of neural channel estimators under unseen Doppler conditions. Motivated by Wiener filter analysis, the approach uses a flat Doppler spectrum whose support covers the expected Doppler range. The proposed criteria impose no specific architectural constraints and can be applied to different neural channel estimators.

Critical strengths

Simple and Practical Design The approach modifies only the synthetic training data distribution, without requiring architectural changes or online adaptation.

Insightful Analytical Justification Through Wiener filter analysis, the paper derives conditions under which a flat Doppler spectrum makes estimation error insensitive to changes in the target Doppler spectrum, justifying the proposed design.

Consistent Empirical Results Experiments on InterpolateNet and Channelformer demonstrate the effectiveness of the approach across two neural architectures.

Critical weaknesses

Incremental Contribution. The paper addresses Doppler mismatch and evaluates its compatibility with the previous PDP design. However, the central prescription closely resembles the earlier framework. While the temporal extension is useful, the conceptual advance over the previous design principle seems to be limited.

Limited Evaluation. Both training and evaluation of the proposed design are conducted at a fixed high SNR of 20dB. Although this provides a controlled setting for evaluating Doppler mismatch, it remains unclear whether the observed robustness persists when the SNR differs.

Unclear Advantage of Approach. In Fig. 3(c),(d), Rounded-spectrum training appears to achieve lower NMSE than flat-spectrum training across all five evaluated spectra for both architectures. Although the authors explicitly prioritize variation across conditions over absolute error, smaller variation alone does not establish a practical robustness advantage.

Insufficient Analysis of Performance Degradation. Fig. 1 shows degradation on Gaussian and BiGaussian spectra. Although the authors acknowledge these exceptions, they do not explain their cause. The paper should explain what the observed degradation implies for the method’s applicability.

Questions for authors

See Weaknesses

Reviewer_nxJgOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

Neural channel estimators for OFDM often get worse on channels they weren't trained on, especially at high speed where Doppler effects are large. Online fine-tuning adds latency and compute. It also can't supply full ground-truth channel labels, so it loses the networks' interpolation advantage. The paper gives a rule for generating synthetic training data: train with a flat (uniform) Doppler spectrum at a maximum Doppler shift at least as large as any shift expected in deployment. Because the data is synthetic, the networks can be trained on complete ground-truth channel matrices, offline only, with no knowledge of the deployment channel. The validation is weak in many ways: Weak baseline. The MMSE comparison works in frequency only, then interpolates bilinearly in time. That is exactly the dimension under study. Even so, MMSE beats the networks by about 30× at low Doppler. There's no 2-D LMMSE and no robust LMMSE (Li et al. 1998), which needs the same priors as the proposed rule. No real alternatives tested. The rule is never compared with randomized-Doppler training or with diverse sampling from standard 3GPP TDL/CDL models. Without that, nothing shows the rule is needed. Only the authors' own networks. There are no third-party neural baselines and no capacity sweep, so "independent of architecture" rests on two hand-picked models. NMSE only. There's no BLER or throughput with coding, so the effect on link performance is unknown. Narrow conditions: One SNR (20 dB) for training and testing. No real high-speed-train profile, despite citing it as motivation. Possibly no inter-carrier interference: eq. 5 assumes the channel is constant within a symbol. Doppler studied alone. Delay generalization is checked only at 70–100 km/h, so delay and Doppler are never varied together at high speed.

Critical strengths

The research direction is relevant and timely. Neural channel estimators often degrade on unseen channels, and online fine-tuning is limited by latency, on-device compute, and the lack of full ground-truth labels. Designing offline training data so the trained models generalize without adaptation is therefore a practical problem worth studying, particularly for high-mobility scenarios.

Critical weaknesses
  1. The contribution isn't shown to be new or necessary. The core theory says that designing for a flat Doppler spectrum at the maximum shift makes the MSE independent of the true spectrum shape. That restates the robust-LMMSE result of Li, Cimini & Sollenberger (1998). The frequency-domain counterpart is the authors' own earlier work (TCCN 2025). In practice the rule amounts to drawing each path's Doppler uniformly from [−f_max, f_max], which standard channel models (3GPP 38.901 TDL/CDL) already make easy. The paper never compares against either obvious alternative:

a 2-D robust LMMSE that uses the same design and needs the same priors; training on diverse data sampled from standard models. Even the weak frequency-only MMSE baseline beats the networks by about 30× at low Doppler.

  1. The validation can't support the claim of practical robustness.

The evaluation reports only estimator NMSE: no BLER or throughput, so the effect on the link is unknown. Only Doppler is varied, on SISO Rayleigh channels at a single SNR (20 dB). Delay and angles aren't varied together with Doppler; the delay check runs only at 70–100 km/h. Inter-carrier interference appears to be ignored, which makes the "very high Doppler" results optimistic. Only the authors' own two architectures are tested. The paper calls the Gaussian and BiGaussian results "slight variation", yet they are about 4–5× worse, and it doesn't explain them.

Questions for authors

Comparison with a strong classical estimator. How do the networks compare with a 2-D time–frequency robust LMMSE built on the same flat-Doppler, maximum-delay design? It needs the same prior knowledge as your rule. The current MMSE baseline works in frequency only and interpolates bilinearly in time, yet still beats the networks by about 30× at low Doppler. Please report NMSE and complexity (FLOPs or latency). If the robust LMMSE matches the networks, the case for a neural estimator weakens a lot. Comparison with existing data generation. Does training on the flat spectrum beat diverse sampling from standard models? Examples: 3GPP 38.901 TDL/CDL with randomized speed and delay spread, or Jakes with f_d drawn uniformly from [0, f_max]. Without this comparison, the need for the proposed rule isn't established. System-level performance. Do the NMSE results carry over to BLER or throughput? For example, NR LDPC at several MCS levels, against LS, LMMSE and genie-CSI receivers. Inter-carrier interference. Is ICI simulated? Eq. 5 assumes the channel is constant within an OFDM symbol. At 1135 Hz with 30 kHz spacing, ICI is about −23 dB, comparable to the 20 dB SNR, so leaving it out would make the high-Doppler results optimistic. The Gaussian and BiGaussian results. NMSE is about 4–5× worse (0.028 vs 0.006). Is this because the Gaussian spectrum's unbounded support puts about 16% of its power beyond ±f_d, breaking your condition? Please define the BiGaussian spectrum. Numbers that need clarification: Sec. 4.1 gives identical NMSE increases (0.0226 and 0.0088) for both networks. Is that a copy-paste error? LS NMSE of about 0.39 at 20 dB SNR looks very high. Why? Is the tested Doppler range 0–350 Hz or 0–1135 Hz? Other architectures. Does the rule work for estimators beyond your own two networks, such as ChannelNet, ReEsNet or DeepRx? Right now the claim that it is "independent of architecture" rests on two models from the same group.

Reviewer_xNn3OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This manuscript works on the robust channel estimation in wireless communication, with the objective to synthesize data for training network for channel estimation across wireless channels, especially for very high Doppler shifts and different Doppler spectra. The motivation is reasonable, and the task of robust channel estimation is important and fundamental in wireless communication. The paper presented details on the backgrounds of system configuration, channel estimation. Then it presents design criteria for Doppler. With the synthesized dataset, it demonstrates better robustness to estimate unseen channels for large range of Doppler shifts.

Critical strengths
  1. The task of robust channel estimation is a fundamentally important task in wireless communication. The neural network-based methods have been widely applied in this area.

  2. The proposed method concentrates on the training data synthesis considering the doppler design, and the data is synthesized with solid physical mechanism.

  3. The results demonstrate that the synthesized training data facilitate better generalization of trained network to different Doppler shifts and spectra.

Critical weaknesses

I have serval concerns on the presentation, novelty and experimental evaluations.

  1. The methodology section presented the backgrounds of system configuration, propagation channels, channel estimation methods and the design criteria for Doppler. These details are necessary, but it is unclear which parts are the novel contributions of this work. Please clarify what is the novel contribution in the synthesizing and what is the key designs that benefits the following robust network training.

  2. The paper lacks more detailed presentation of the training/test dataset construction process, network structures and training details. The paper experimented with different training/test Doppler shifts for training and evaluating neural networks. Though it demonstrates some robustness of the learned InterpolateNet and Channelformer, it did not compare with any other related methods for channel estimation. Moreover, the different SNR is also a factor for testing generalization.

  3. Though ablation study is presented in section 4.3 on the design criteria, it is mainly for evaluation of violating the design. It should include more analysis on the benefits of the design criteria itself. For example, what specific design contributes best to the robustness of trained networks?

  4. Overall, this is an application work for channel estimation in wireless communication. Though the method concentrates on the machine learning approach in this area, it lacks significant contributions to the learning approach itself. The data synthesis criteria is the major contributions, however, it lacks clear novelty in this design and deeper understanding on how it benefits the downstream neural network training.

Questions for authors

As discussed above, I have several concerns on the unclear novelty, limited contributions to learning approach, insufficient evaluations and ablation studies to analyze the benefits of the design criteria. This is a machine learning-based research for channel estimation, it should demonstrate sufficient methodology contributions to meet the bar of ICLR.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.