acceptodds
Under review as a conference paper at ICLR 2027

Gradient-Guided Physics-Informed Multi-Fidelity U-Mamba for Discontinuous Flow Field Prediction

Abstract

Surrogate modeling of strong discontinuities, such as shock waves and contact discontinuities, in compressible flows remains challenging: high-resolution numerical schemes are computationally prohibitive on fine grids, purely data-driven operators lack physical consistency and induce Gibbs oscillations, and standard physics-informed neural networks (PINNs) break down where strong-form differential residuals become ill-defined. We propose PIMUG, a Gradient-Guided Physics-Informed Multi-Fidelity U-Mamba framework coupling three complementary branches: (1) a multi-fidelity trunk extracting spatio-temporal features from multi-resolution, multi-timestep inputs; (2) an explicit gradient branch supervising directional derivatives to sharpen discontinuity fronts without oversmoothing; and (3) a coordinate-conditioned physics branch with a divergence-adaptive residual loss that enforces Euler conservation while relaxing ill-posed penalties at shock fronts. An Input Compression Module (ICM) suppresses spatio-temporal redundancy across fidelity levels. On 2D Riemann benchmarks spanning slip lines, rarefactions, and shock collisions, PIMUG preserves sharp discontinuity profiles without visible Gibbs oscillations. As a multi-fidelity field corrector, PIMUG consumes low-fidelity numerical solutions at the target time—a richer input regime than instance-wise pointwise PINNs (PINN-WE), which solve from initial and boundary conditions alone; under this setting, PIMUG reduces self-enforced Euler residuals by two to four orders of magnitude relative to PINN-WE with millisecond amortized inference. Compared to Fourier Neural Operators (FNO) on identical inputs, PIMUG eliminates Gibbs blurring while retaining physical consistency; augmenting FNO with the same residual loss merely induces degenerate over-smoothing while degrading data accuracy by two orders of magnitude, isolating PIMUG's gains as architectural rather than a loss artifact.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
physics-informed neural networks, multi-fidelity surrogate, discontinuous flows, Mamba state space models, Riemann problem, Euler equations, gradient guidance

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_9pt2OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

Surrogate modeling of strong discontinuities in compressible flows presents significant challenges, as traditional data-driven models induce Gibbs oscillations and standard physics-informed neural networks (PINNs) fail where differential residuals are ill-defined. The authors propose PIMUG, a Gradient-Guided Physics-Informed Multi-Fidelity U-Mamba framework that integrates multi-fidelity feature extraction, explicit gradient supervision to sharpen fronts, and a divergence-adaptive residual loss.

Critical strengths
  • The research addresses the highly complex and novel problem of modeling strong discontinuities like shock waves in compressible flows, specifically navigating the breakdown of standard PINNs in regions where strong-form differential residuals become ill-defined.

  • PIMUG achieves millisecond amortized inference while reducing self-enforced Euler residuals by two to four orders of magnitude.

Critical weaknesses
  • [1] also focuses on shock waves and sharp discontinuities, representing an important related work that is neither compared nor discussed.

  • The authors highlight multiple times in the abstract and introduction how their proposed method reduces Gibbs oscillations and blurring, but it is unclear which experimental metric justifies this point.

  • While the inclusion of neural operators and spatial-temporal modeling techniques in related work is appreciated, the related work discussion mainly focuses on very old papers and should include more recent ones especially on neural operators.

  • The baseline comparison is insufficient, relying only on a PINN variant and FNO (an older model with many recent superiors); more recent baselines should be included.

  • The proposed model relies on a U-Mamba backbone and existing building blocks, presenting no major breakthrough in neural modeling.

  • The experiments are only conducted on four 2D Riemann cases, leaving it unclear whether the model scales or can be applied to 3D data and irregular meshes.

Questions for authors
  • Are the experiments evaluating rollout or next-step predictions?

  • Can the authors provide more details on the dataset, specifically whether it is self-generated or from existing works, what solvers were used, the number of trajectories and frames, and the train/test split?

Reviewer_56zuOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes PIMUG, a multi-fidelity U-Mamba framework for reconstructing discontinuous compressible flow fields. The architecture integrates a multi-fidelity backbone, a gradient-guidance branch, coordinate-conditioned physics regularization, and an input compression module. Experiments on 2D Riemann problems evaluate reconstruction accuracy, physical consistency, discontinuity preservation, and cross-case transfer.

Critical strengths
  1. The paper addresses an important problem in scientific machine learning, particularly the preservation of shocks and contact discontinuities.

  2. Combining multi-fidelity information with explicit gradient supervision is technically interesting and potentially useful for reconstructing sharp flow structures.

  3. The paper provides detailed methodological descriptions, ablation studies, and a cross-case transfer experiment. The authors also acknowledge several limitations in their evaluation.

Critical weaknesses
  1. Questionable physical consistency. The continuous Euler residual is computed through the coordinate-conditioning pathway rather than the complete derivatives of the reconstructed flow. Therefore, its small magnitude does not establish actual conservation. PIMUG also exhibits higher discrete residuals than FNO in Cases 1 and 2.

  2. Insufficient quantitative advantages. FNO achieves substantially lower reconstruction MSE across all three benchmark cases. At the +5% perturbation, PIMUG's errors are approximately 17, 132, and 14 times higher. The claimed advantages in discontinuity preservation lack sufficient quantitative support, such as shock-position errors, transition widths, and local discontinuity errors.

  3. Limited architectural novelty and incomplete ablations. The framework combines several established components. Existing ablations do not systematically isolate their contributions within the complete architecture. Parameter-matched comparisons are also missing.

  4. Insufficiently controlled baseline comparison. Physics-augmented FNO uses a discrete Euler residual, whereas PIMUG uses a coordinate-path automatic-differentiation residual. These objectives are not equivalent. The observed degradation of FNO may also reflect inappropriate loss weighting.

  5. Limited practical validation. The method performs same-time multi-fidelity reconstruction rather than genuine future-state forecasting. Its inference-time evaluation excludes low-fidelity simulation costs. Cross-case transfer significantly degrades physical consistency, and single-run results lack statistical reliability.

Questions for authors
  1. How can the authors demonstrate that the coordinate-conditioned physics branch improves actual conservation of the reconstructed flow fields?

  2. Can the authors provide quantitative evidence of superior discontinuity preservation despite the substantially higher MSE compared with FNO?

  3. What are the individual contributions of each component under controlled, parameter-matched ablations?

  4. How sensitive are the physics-augmented FNO results to the residual-loss weight?

  5. What is the end-to-end computational cost when low-fidelity numerical simulations are included?

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

This paper reconstructs high-precision solutions at target time steps using low-precision flow fields from two resolutions and two adjacent time instances, combining U-Mamba, gradient branch, coordinate condition branch, and input compression modules, validated on three types of 2D Riemann problems. Ablation studies support the effectiveness of the combined design, but it remains unclear whether the relative advantage over FNO indicates improved conservation properties, or whether reduced continuous residuals necessarily imply better physical consistency.

Critical strengths

Inputs at two resolutions provide information about discontinuities with varying degrees of smoothness, while the gradient branch directly leverages changes near discontinuities. In Table 7, the full model achieves lower data errors than all ablated variants across six test perturbations, supporting the practicality of the integrated design—performance gains are not solely due to replacing the backbone.

The authors explicitly report differences in input configurations compared to PINN-WE, lower data errors than FNO, and increased residual magnitudes after freezing the encoder for transfer learning, which helps clarify applicability and highlights that visual similarity between reconstructed and reference flow fields does not guarantee satisfaction of physical constraints.

Critical weaknesses

PIMUG fixes input flow fields and performs differentiation only along the coordinate branch, so the resulting continuous residual does not represent the actual spatiotemporal residual of the reconstructed flow field. The output's insensitivity to coordinates may also lead to small residual values. Removing adaptive weighting fails to eliminate distinctions from PINN-WE, meaning a reduction in residuals by two to four orders of magnitude is insufficient to prove corresponding improvements in conservation.

Under identical inputs, PIMUG exhibits data errors approximately 17 times and 132 times higher than FNO in Case 1 and Case 2, respectively, with similarly elevated gradient errors. Current figures do not clearly illustrate sufficient improvements to justify these discrepancies. Since the residual operators and gradient objectives differ between physics-enhanced FNO and PIMUG, poor performance of the former cannot independently validate architectural superiority.

Questions for authors

How does the current formulation distinguish between outputs satisfying the Euler equations and those merely being insensitive to the coordinate branch? Please explain, based on existing implementation, the relationship between the continuous residual and actual flow field conservation.

Based on available profiles and error results, please identify the most definitive improvement in discontinuity reconstruction compared to pure-data FNO. Given the differing residual operators and gradient objectives, to what extent can the results of physics-enhanced FNO actually attribute performance gains to architectural advantages?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.