acceptodds
Desk rejected at ICLR 2027

Moving Receiver Signal Prediction via Wavefield–Trajectory Decoupling

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

Predicting future electromagnetic signals at moving receivers is an important task in mobile wireless systems, supporting beam tracking, Doppler compensation, and resource allocation. The received signal varies both as the wavefield evolves over time and as receiver motion changes the sampled field amplitude and phase. In particular, motion-induced Doppler shifts can produce rapid phase variations, making accurate prediction challenging. Existing methods tend to jointly learn temporal wavefield evolution and spatial trajectory sampling to predict future signals. Encoding these two factors with different properties into a shared representation can cause different receiver velocities to induce conflicting gradient directions during optimization, thereby impairing the learned wavefield representation. To focus the network on wavefield evolution while avoiding neural approximation of trajectory sampling, we propose Wavefield–Trajectory Decoupling (WTD), comprising a velocity-agnostic complex neural operator and a parameter-free geometric projector. The operator predicts Eulerian wavefields from past wavefields, and the projector samples the predicted field along known trajectories to obtain future received signals. A phase-preserving complex activation and full-field Sobolev supervision constrain high-frequency wavefield learning. WTD outperforms baselines on in-distribution and out-of-distribution velocity, with up to lower relative signal error and lower absolute Doppler error.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Neural operators, Electromagnetic wavefields, Signal prediction, Velocity generalization

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_CW7zOpenReview ↗
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
AuthorsOpenReview ↗

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:

  1. 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.

  2. 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.

  3. 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.

  4. 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.

  5. 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.

AuthorsOpenReview ↗

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 s[n]=u(rn,tn)s[n]=u(r_n,t_n) 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.

AuthorsOpenReview ↗

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 0.10c0.10c 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.

ConditionBias scaleSignal Rel-L2 ↓Doppler error ↓Frame-to-frame phase increment error (°) ↓Field N-L2 ↓
Exact trajectory00.10110.013131.6030.0972
Along-track position bias0.02λc0.02\lambda_c0.19390.017101.9860.0972
Along-track position bias0.05λc0.05\lambda_c0.37120.016482.6080.0972
Along-track position bias0.10λc0.10\lambda_c0.67260.012093.4170.0972
Transverse position bias0.02λc0.02\lambda_c0.10280.013001.5610.0972
Transverse position bias0.05λc0.05\lambda_c0.11400.012471.6530.0972
Transverse position bias0.10λc0.10\lambda_c0.14440.012322.0310.0972
Along-track velocity bias0.02λc0.02\lambda_c0.14520.015271.8360.0972
Along-track velocity bias0.05λc0.05\lambda_c0.24120.021302.2930.0972
Along-track velocity bias0.10λc0.10\lambda_c0.41060.035923.3080.0972
Transverse velocity bias0.02λc0.02\lambda_c0.10160.012951.5560.0972
Transverse velocity bias0.05λc0.05\lambda_c0.10640.012741.5600.0972
Transverse velocity bias0.10λc0.10\lambda_c0.12180.012531.7130.0972

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 conditionSignal Rel-L2 ↓Doppler error ↓Frame-to-frame phase increment error (°) ↓Field N-L2 ↓
Along-track position bias0.3712 → 0.10190.01648 → 0.012732.608 → 1.5850.0972 → 0.0972
Transverse position bias0.1140 → 0.10120.01247 → 0.013131.653 → 1.5980.0972 → 0.0972
Along-track velocity bias0.2412 → 0.10050.02130 → 0.012372.293 → 1.5570.0972 → 0.0972
Transverse velocity bias0.1064 → 0.10220.01274 → 0.013031.560 → 1.6040.0972 → 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.

Reviewer_mi1hOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

In this paper, the authors propose Wavefield–Trajectory Decoupling (WTD), a framework for predicting received signals at moving receivers. The proposed approach aims to separate electromagnetic wavefield evolution from receiver motion, rather than directly incorporating receiver velocity into signal prediction. Specifically, WTD uses a velocity-independent neural operator to predict future Eulerian wavefields from past observations, while a trajectory-based projector samples the predicted fields along known receiver trajectories to obtain future received signals.

Critical strengths

A key strength of this paper is its clear separation between what needs to be learned by the neural network and what can be handled using known physical principles. Unlike conventional velocity-conditioned approaches, such as FiLM, AdaLN, or feature concatenation, which require the network to jointly learn wavefield evolution and its dependence on receiver motion, WTD focuses on predicting the scene-dependent wavefield. The effects of receiver motion, including Doppler shifts, are then captured through parameter-free geometric sampling along known trajectories. This provides a well-motivated inductive bias that reduces the complexity of the learning task and helps explain the improved generalization to unseen receiver velocities.

Critical weaknesses

A major concern is that the experimental setting is highly idealized, which raises questions about the practical applicability of the proposed method. WTD assumes that accurate receiver positions, velocities, and past electromagnetic wavefields are available, which may be difficult to obtain in real-world wireless systems. Moreover, the evaluation is limited to simplified 2D simulated environments with straight-line receiver trajectories, making it unclear how well the method would generalize to realistic mobility scenarios. More importantly, the experiments consider receiver velocities as high as 0.2c. If c represents the speed of light, such velocities are far beyond those encountered in practical wireless communication systems. The authors should clarify the physical meaning of these velocity settings and explain which real-world applications the experimental scenarios are intended to represent.

Questions for authors
  1. The proposed method assumes that past electromagnetic wavefields and accurate receiver positions and velocities are available. Could the authors clarify how such information can be obtained in practical wireless communication systems?

  2. How robust is the proposed method to errors in receiver position and velocity estimation, as well as more complex motion patterns involving acceleration or changes in direction?

  3. The experiments consider normalized receiver velocities up to 0.2c. Does c represent the speed of light? If so, could the authors explain the practical relevance of such high velocities and which real-world applications these settings are intended to represent?

5 comments
AuthorsOpenReview ↗

Official Comment by Authors

We sincerely thank reviewer mi1h for the careful review and for recognizing the clarity of the paper, the physical motivation for decoupling, and the value of its inductive bias. Your comments on input availability, errors in motion estimates, and the practical relevance of the tested speeds helped us clarify the conditions under which WTD applies. We appreciate the opportunity to improve our work based on the reviewer's feedback. Below, we have prepared a detailed response to address the questions raised by the reviewer.

Our response is organized into three key parts:

1. We explain how practical systems could provide the inputs: past wavefields, receiver positions, and receiver velocities. We also describe the steps needed to extend our simulation setup to real measurements.

2. We evaluate trajectory errors and complex motion using three frozen checkpoints and all 60 test fields. We quantify the effects of position and velocity biases and errors from trajectory extrapolation. We also test prediction along given accelerating and turning trajectories and evaluate whether past fields and received signals available before prediction can help correct fixed biases.

3. We clarify the meaning and practical relevance of the tested speeds, including the background wave speed cc, the normalized receiver speed v/cv/c, and the prediction horizon. We explain why the controlled numerical tests extend to 0.2c0.2c and what comparing displacements in units of wavelength can tell us about practical scenarios.

We provide the experimental results and explanations for each point below.

AuthorsOpenReview ↗

Official Comment by Authors

[W1, Q1] 1. Input sources and practical requirements

We thank the reviewer for the careful review and for recognizing the paper's clarity, the physical motivation for decoupling, and the role of this inductive bias in reducing learning complexity and improving generalization to unseen velocities. You identified the central idea of WTD: separating wavefield evolution, which the neural network learns, from trajectory sampling, which follows known physical relationships. WTD learns a shared, velocity-independent wavefield and uses parameter-free geometric projection to generate signals for receivers following different paths. The sampling process naturally captures motion-induced phase changes and Doppler effects.

Section 3.1 defines an offline electromagnetic surrogate prediction setting. Using the same past fields and motion information for the methods in this comparison lets us assess how the architectural design affects velocity extrapolation.

Validation with real measurements requires coherent spatial observations, field estimation, antenna coordinate registration, and time synchronization, as well as separate assessment of the errors in each step. The original manuscript tested the mechanism under controlled conditions and gave limited detail on how these inputs could be obtained in practice. Your question helps us specify the input requirements and clarify how the design could be used in practical systems.

Measurement, processing, and feedback delays create a need for signal prediction in mobile wireless systems. We aim to provide future received-signal information for beam tracking, Doppler compensation, and resource allocation. In one possible deployment, the wireless system supplies the scene wavefield and the mobile platform supplies its motion state.

Past electromagnetic fields. A system with a scene propagation model could use coherent channel sounding at known positions to calibrate its environmental and propagation parameters. With the base station's known transmit waveform and the antenna responses, the system could then estimate the past complex field in a local region around the receiver and supply it to WTD. WTD predicts the future wavefield and samples it along the receiver path to generate the future received signal. In this setup, the system handles spatial measurements and field estimation, and the terminal supplies its motion information. Our experiments use solver-generated past fields to evaluate prediction and projection once those fields are available.

Receiver position. Mobile platforms such as trains and UAVs can estimate their current position by fusing GNSS and inertial navigation. The system can then register the receiving antenna position to the wavefield coordinates and align its timestamps with the field snapshots to locate the field patch and determine the starting point for projection.

Receiver velocity. As described in Section 3.1, onboard navigation or platform telemetry can provide a velocity vector, for example by fusing GNSS Doppler or time-differenced carrier-phase observations with inertial measurements. Sampling positions can then be extrapolated over a local window under the approximately constant-velocity assumption.

Motion estimates from practical systems have finite accuracy. In Q2, we quantify the effects of position and velocity biases separately and test whether waveforms received before the forecast can help correct fixed biases when accurate, synchronized past fields are available. We will clarify these input sources and requirements to connect the wireless application motivation with the offline conditions evaluated in this work.

AuthorsOpenReview ↗

Official Comment by Authors

[W2, Q2] 2. Trajectory errors and complex motion

We conducted supplementary evaluations on all 60 test fields using three WTD checkpoints from the main experiments. For each field and checkpoint, all conditions use the same complete past field and predicted wavefield. We vary only the coordinates supplied to the geometric projector. We report Signal Rel-L2, carrier-normalized Doppler error, amplitude-weighted frame-to-frame phase increment error, and Field N-L2.

True motion and supplied trajectorySignal Rel-L2 ↓Doppler error ↓Phase increment error ↓Field N-L2 ↓
Straight line at constant speed, exact trajectory0.10060.01281.84°0.0972
Straight-line acceleration, exact trajectory0.10340.01322.13°0.0972
45° turn, exact trajectory0.09410.01011.69°0.0972
Same accelerating path, constant-velocity extrapolation0.26280.03132.98°0.0972
Same turning path, extrapolation along the initial tangent0.15490.01762.33°0.0972

With exact trajectories, acceleration and turning yield errors close to those for straight-line motion at constant speed. Using constant-velocity extrapolation for the same true paths increases all three signal errors. As the extrapolated coordinates diverge from the receiver's true positions, the projector samples the predicted field at the wrong locations. Spatial variations in the field then translate into errors in signal amplitude, phase, and phase evolution, affecting both the complex waveform and Doppler estimates. These comparisons quantify the cost of trajectory extrapolation that misses a maneuver. Field N-L2 stays the same because each trajectory uses the same predicted field.

Fixed position and velocity biases also shift the sampling coordinates. We therefore added calibration based on agreement with past received waveforms. We fit position and velocity corrections by matching the waveform sampled along a candidate path in the past field to the waveform received before the forecast. We then apply the corrections to future sampling coordinates. We fit on the first 30 history frames and check the fit on the remaining 10, accepting a correction only if it reduces the matching error on those held-out frames. The calibrator uses only pre-forecast data, with no ground-truth position references or future received signals. It addresses fixed biases. Acceleration or turning that first occurs during the forecast is absent from the history used for calibration. The arrows compare errors before and after calibration under the same bias condition.

Trajectory conditionSignal Rel-L2 ↓Doppler error ↓Phase increment error ↓Field N-L2 ↓
Exact-trajectory control0.1011 → 0.10110.0131 → 0.01311.60° → 1.60°0.0972
Longitudinal position bias0.3712 → 0.10190.0165 → 0.01272.61° → 1.59°0.0972
Longitudinal velocity bias0.2412 → 0.10050.0213 → 0.01242.29° → 1.56°0.0972
Transverse position bias0.1140 → 0.10120.0125 → 0.01311.65° → 1.60°0.0972
Transverse velocity bias0.1064 → 0.10220.0127 → 0.01301.56° → 1.60°0.0972

Calibration reduces all three signal errors for both longitudinal biases. This shows that waveforms received before the forecast can help mitigate fixed position and velocity biases when complete, synchronized simulated past fields are available. For the transverse biases, Signal Rel-L2 decreases slightly, while Doppler error increases. The benefit therefore depends on the bias direction and the metric. We use the same WTD weights and predicted field before and after calibration. The improvement comes from correcting the sampling trajectory, and Field N-L2 stays the same.

Reviewer_RvA2OpenReview ↗
Rating
1
Confidence
1
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

I believe the topic of this paper falls outside the scope of ICLR. Therefore, I recommend a desk rejection without proceeding to a full review.

Critical strengths

Not Applicable.

Critical weaknesses

I believe the topic of this paper falls outside the scope of ICLR. Therefore, I recommend a desk rejection without proceeding to a full review.

Questions for authors

Please refer to the weakness section.

1 comment
AuthorsOpenReview ↗

Official Comment by Authors

Dear Reviewer RvA2,

Thank you for raising the question of fit with ICLR. The ICLR 2027 Call for Papers explicitly includes applications to physical sciences. Our submission uses moving-receiver signal prediction to study a machine learning question in that area: how to design a neural operator that generalizes when the underlying field dynamics are shared across conditions but the observations depend on the receiver trajectory.

Our contribution is to separate learned field evolution from trajectory-dependent sampling. WTD learns a shared, velocity-independent wavefield predictor and uses a parameter-free geometric projector to generate signals along supplied trajectories. Table 1 evaluates this design at receiver speeds beyond the training range. Section 4.4 further compares WTD and FiLM with and without future-field supervision. Under signal-only supervision, WTD achieves about 38% lower OOD Doppler error than FiLM in the fixed-seed comparison in Table 10.

We would appreciate your reconsideration of the scope assessment in light of this connection to ICLR’s stated scope and the paper’s contribution to neural operator design and generalization. We would welcome any specific questions about the method or evidence.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Request for AC Attention Regarding Reviewer RvA2's Non-Substantive Review (Confidence: 1)

Dear Area Chair,

Thank you very much for your time and effort in managing our submission (#42555).

We would like to respectfully bring to your attention the review provided by Reviewer RvA2 (Rating: 1, Confidence: 1). The review contains a single sentence recommending a desk rejection on the grounds that the paper "falls outside the scope of ICLR," without summarizing the paper or providing any technical evaluation.

We respectfully note that our work squarely fits ICLR’s scope under our selected primary area of Applications to Physical Sciences (AI for Science):

  • Core Focus on Learning Representations: Our paper addresses a fundamental representation learning and OOD generalization problem in Neural Operators. Specifically, we show that encoding physical wavefield evolution and observer motion into a shared conditioned representation causes gradient conflicts, and we propose a decoupled neural operator to learn velocity-invariant representations.
  • Consensus from Knowledgeable Reviewers: Both Reviewer mi1h and Reviewer CW7z (Confidence: 3) clearly understood the problem setting and positively recognized our architectural inductive bias and neural operator design.

Given that Reviewer RvA2 explicitly stated "I do not have sufficient expertise to assess this submission reliably and have alerted the AC" (Confidence: 1) and did not provide a substantive review, we kindly ask that you take this into consideration when weighing the reviews.

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

Related papers

Discussion (0)

Sign in to comment.