- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
The main strength of the paper is the clear architectural principle of separating wavefield evolution from trajectory sampling. Since receiver velocity affects where the field is sampled rather than how the field itself evolves, the proposed decoupling is physically meaningful and better motivated than simply injecting velocity into a learned predictor. The parameter-free geometric projector is also a clean design choice that reduces the burden on the neural network. The empirical results are strong within the studied setting. WTD substantially outperforms FiLM, concatenation, adaLN, and FNO baselines in both ID and OOD velocity regimes. The ablations on field supervision, gradient loss, complex activation, and velocity handling support the importance of the proposed components. The additional MEEP evaluation is also useful because it reduces the concern that the results are specific to one custom FDTD solver.
Critical strengths
The main strength of the paper is the clear architectural principle of separating wavefield evolution from trajectory sampling. Since receiver velocity affects where the field is sampled rather than how the field itself evolves, the proposed decoupling is physically meaningful and better motivated than simply injecting velocity into a learned predictor. The parameter-free geometric projector is also a clean design choice that reduces the burden on the neural network. The empirical results are strong within the studied setting. WTD substantially outperforms FiLM, concatenation, adaLN, and FNO baselines in both ID and OOD velocity regimes. The ablations on field supervision, gradient loss, complex activation, and velocity handling support the importance of the proposed components. The additional MEEP evaluation is also useful because it reduces the concern that the results are specific to one custom FDTD solver.
Critical weaknesses
The main weakness is that the task setting may be too restrictive compared with practical moving-receiver signal prediction. At inference, the method assumes access to local past 2D wavefield patches and known receiver velocity. In many wireless sensing or communication scenarios, the receiver may only observe its own 1D signal history, rather than the full local Eulerian field. This assumption is central to WTD's advantage, but the paper does not sufficiently discuss how such 2D field inputs would be obtained in practice or how robust the method is when the field observation is incomplete, noisy, or estimated. A second concern is that the comparison may partly favor WTD because it receives dense future-field supervision during training, while the practical target is the 1D received signal. Although the authors include FiLM+field and ablation studies, the large performance gap raises the question of whether the improvement mainly comes from the availability of full-field supervision and strong physical structure, rather than from the decoupling principle alone. More matched comparisons under limited or signal-only supervision would make the conclusion stronger. Finally, the generality of the method is not fully established. The experiments are based on simulated 2D TM wavefields, fixed local windows, approximately constant receiver velocity, and a limited set of speed values. It remains unclear how well the method extends to more realistic 3D propagation, complex multipath environments, acceleration, unknown or noisy trajectories, different antenna/receiver models, and real measurement data.
Questions for authors
In practical wireless systems, how would the local 2D past wavefield patch be obtained at inference time? Is WTD still applicable if only the receiver's 1D signal history is available?
How sensitive is WTD to errors in the known receiver velocity or initial position? Since the geometric projector directly uses these quantities, even small trajectory errors may lead to phase errors.
Can the authors provide a more controlled comparison where all methods receive the same level of field supervision, or where WTD is trained with only signal-level supervision?
Does the method extend to accelerated receiver motion or non-linear trajectories? The current formulation assumes approximately constant velocity within each local prediction window.
Can the authors discuss how the method would scale to 3D electromagnetic fields or real measured data?
7 comments
Official Comment by Authors
[W1, Q1] 1. Past wavefield inputs and 1D observations
Thank you for pointing out this key input requirement. We use local past wavefields to give the field operator information about spatial amplitude, phase, and how they change over time. The operator can then focus on predicting how the wavefield evolves in the scene, and the geometric projector generates the received signal by sampling the predicted field along the receiver trajectory. This separates what the network needs to learn from what can be computed using known physical relationships, reducing the burden of learning field evolution and motion-dependent sampling jointly. To test this design, WTD and the velocity-conditioned baselines in our main experiments use the same 2D past wavefields. The comparison evaluates whether explicit decoupling improves prediction across velocities when this information is available.
Obtaining local past fields in practice. Our experiments use local past fields to test the decoupling principle. A practical system would need a source of this spatial information. One possible approach is for the system to collect phase-coherent complex channel observations at known positions and use them to calibrate a propagation model of the scene. Given the transmit waveform and antenna responses, the system could then estimate the past complex field around the receiver and extract a local patch using common spatial, temporal, and phase references. The terminal can retain its local 1D observations while the system supplies the spatial field information WTD needs. If the system can provide suitable past fields, WTD can make predictions even when the terminal itself records only a 1D signal. Our experiments use solver-generated past fields to evaluate the prediction and projection stages. The accuracy of field estimation and the quality of synchronization in the proposed measurement pipeline still need to be evaluated.
Applicability when only 1D signal history is available. The question is whether the system has enough spatial information to sample the field along the future trajectory. Our target is also a 1D received signal. WTD predicts it in two steps by forecasting a shared wavefield and sampling it along the trajectory. In our observation model, the 1D history satisfies and records field values along the past trajectory. A local 2D history also provides the amplitude, phase, and temporal variation of the field in the surrounding region. This gives the operator information about locations that the receiver may visit later and allows the same field prediction to serve different trajectories. The decoupled method evaluated here therefore requires a 2D past field as input.
If the system can estimate the past field from a scene model or spatial observations, WTD can be used even when the terminal records only a 1D signal. If the 1D history is the only information available to the entire system, applying WTD would require an additional field estimation step that uses a scene model or propagation priors. Our current results do not cover that setting. We will clarify this input requirement and emphasize that our main experiments demonstrate the value of explicit trajectory sampling for prediction across velocities when the compared methods share the same 2D past fields.
Official Comment by Authors
[W2, Q2] 2. Sensitivity to position and velocity errors
Thank you for highlighting how motion-estimation errors affect geometric projection.
Receiver position. Mobile platforms such as trains and UAVs can estimate their position by fusing GNSS and inertial navigation. The system registers the receiving antenna position to the wavefield coordinate frame and aligns its timestamp with the field snapshots to locate the local patch and set the projection starting point.
Receiver velocity. As described in Section 3.1, onboard navigation or platform telemetry can provide a velocity vector. The navigation system can estimate it from GNSS Doppler or time-differenced carrier-phase (TDCP) observations fused with inertial measurements. Over a short window, the system can extrapolate sampling positions under a locally constant velocity assumption.
A position bias shifts the sampling locations, while a velocity bias produces a displacement that grows over the forecast.
We tested deterministic biases at a reference speed of on 60 test fields with three frozen checkpoints. We held the past fields, reference signals, and model weights fixed and varied only the coordinates supplied to the projector. Position biases are fixed offsets. For velocity biases, the initial position is exact and the bias scale is the displacement at the last forecast frame. Both scales are expressed in carrier wavelengths. Each bias row averages both signs across fields and checkpoints. Doppler error is carrier-normalized and uses the archived lag-1 estimator with an additional magnitude weight on each lag-1 product. The phase increment error is amplitude-weighted.
| Condition | Bias scale | Signal Rel-L2 ↓ | Doppler error ↓ | Frame-to-frame phase increment error (°) ↓ | Field N-L2 ↓ |
|---|---|---|---|---|---|
| Exact trajectory | 0 | 0.1011 | 0.01313 | 1.603 | 0.0972 |
| Along-track position bias | 0.1939 | 0.01710 | 1.986 | 0.0972 | |
| Along-track position bias | 0.3712 | 0.01648 | 2.608 | 0.0972 | |
| Along-track position bias | 0.6726 | 0.01209 | 3.417 | 0.0972 | |
| Transverse position bias | 0.1028 | 0.01300 | 1.561 | 0.0972 | |
| Transverse position bias | 0.1140 | 0.01247 | 1.653 | 0.0972 | |
| Transverse position bias | 0.1444 | 0.01232 | 2.031 | 0.0972 | |
| Along-track velocity bias | 0.1452 | 0.01527 | 1.836 | 0.0972 | |
| Along-track velocity bias | 0.2412 | 0.02130 | 2.293 | 0.0972 | |
| Along-track velocity bias | 0.4106 | 0.03592 | 3.308 | 0.0972 | |
| Transverse velocity bias | 0.1016 | 0.01295 | 1.556 | 0.0972 | |
| Transverse velocity bias | 0.1064 | 0.01274 | 1.560 | 0.0972 | |
| Transverse velocity bias | 0.1218 | 0.01253 | 1.713 | 0.0972 |
Signal Rel-L2 is more sensitive to along-track biases in the tested scenes. This directional difference depends on the local field structure. A fixed spatial offset can produce a nearly constant phase offset, which explains why Signal Rel-L2 can increase with little change in Doppler error. Field N-L2 is unchanged because all trajectories sample the same predicted field for each test field and checkpoint.
We also tested calibration using observations available before the forecast. Given complete, accurate, synchronized past fields, we jointly fit a position correction and a constant velocity correction by matching field samples along the candidate trajectory to the previously received waveform. We fit on the first 30 history frames and accept the correction only if it reduces the matching error on the remaining 10. Calibration keeps the model weights fixed and uses neither ground-truth positions nor future received signals. The arrows compare results before and after calibration for a fixed position offset or final-frame displacement of 0.05 carrier wavelengths.
| Bias condition | Signal Rel-L2 ↓ | Doppler error ↓ | Frame-to-frame phase increment error (°) ↓ | Field N-L2 ↓ |
|---|---|---|---|---|
| Along-track position bias | 0.3712 → 0.1019 | 0.01648 → 0.01273 | 2.608 → 1.585 | 0.0972 → 0.0972 |
| Transverse position bias | 0.1140 → 0.1012 | 0.01247 → 0.01313 | 1.653 → 1.598 | 0.0972 → 0.0972 |
| Along-track velocity bias | 0.2412 → 0.1005 | 0.02130 → 0.01237 | 2.293 → 1.557 | 0.0972 → 0.0972 |
| Transverse velocity bias | 0.1064 → 0.1022 | 0.01274 → 0.01303 | 1.560 → 1.604 | 0.0972 → 0.0972 |
Calibration reduces all three signal errors for both along-track biases, bringing Signal Rel-L2 close to the exact-trajectory baseline. For transverse biases, Signal Rel-L2 also decreases, while Doppler error increases and the phase results are mixed. These results apply to complete, accurate, synchronized simulated past fields and fixed biases that persist across the history and forecast windows.
Official Comment by Authors
We sincerely thank reviewer CW7z for recognizing the value of separating wavefield evolution from trajectory sampling and for noting the strengths of the prediction results, ablation studies, and independent MEEP evaluation. Your comments on input requirements, fairness of supervision, and extensibility helped us clarify what the current experiments establish and what requires further validation.
Our response is organized into five key parts:
Historical wavefield inputs and 1D observations. We describe a possible way for the system to provide local past fields. We distinguish the case where only the terminal has access to its own 1D signal history from the case where that history is the only information available to the entire system, and clarify the inputs that WTD requires.
Sensitivity to position and velocity errors. We evaluate deterministic biases in position and velocity, report four metrics, and test whether previously received waveforms can help correct fixed biases when accurate, synchronized past fields are available.
Contributions of the architecture and field supervision. We present the existing four-way comparison of FiLM and WTD with and without field supervision and use the same checkpoints to report additional signal and phase metrics. We clarify which supervision labels were matched and which training details differed.
Accelerated motion and nonlinear trajectories. We compare straight and curved paths at constant and increasing speeds. The experiments test whether WTD can predict received signals along given per-frame positions without retraining. We clarify the role of externally supplied coordinates and the separate task of estimating future trajectories.
Extension to 3D fields and real measurements. We describe how to extend the field prediction and trajectory sampling steps to three dimensions and address the resulting computational and observation requirements. We also discuss what is needed to obtain fields from measurements, synchronize observations, and calibrate the receiver response.
We provide the results and explanations for each point below.