acceptodds
Withdrawn from ICLR 2027

POD-Flow: A POD-Driven Flow Matching Framework for Physical Field Generation on Irregular Geometries

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Generative models, including diffusion and flow matching models, provide a promising probabilistic framework for solving forward and inverse problems governed by partial differential equations (PDEs), particularly when observations are sparse or noisy. However, most existing approaches rely on U-Net or Fourier-based architectures designed for regular grids, limiting their applicability to physical fields discretized on irregular meshes. We propose POD-Flow, a flow matching framework that incorporates a Proper Orthogonal Decomposition Neural Operator (PODNO) as its generative backbone. The data-derived POD modes form a globally supported basis that preserves the dominant long-range spatial correlations of irregularly discretized physical fields, enabling efficient global modeling without graph message passing or attention mechanisms. POD-Flow adopts direct final-solution prediction rather than conventional velocity prediction, a formulation that naturally aligns with POD projection: the dominant modes retain most of the physical-field energy while filtering out much of the broadly distributed noise energy, yielding high-signal-to-noise estimates early in inference.During inference, observation-guided sampling further steers the generated fields toward available measurements, allowing a single trained model to address forward prediction, field reconstruction, and parameter inversion under different partial-observation settings without retraining. Experiments on multiple PDE benchmarks demonstrate that POD-Flow achieves competitive or improved reconstruction accuracy while requiring substantially fewer sampling steps than diffusion-based baselines. These results establish POD-Flow as an efficient generative framework for probabilistic PDE solving beyond regular-grid discretizations.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Flow Matching, Proper Orthogonal Decomposition, Neural Operator, Irregular Geometries, Partial Differential Equations, Inverse Problems, Sparse Observations

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_Wu2COpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper introduces POD-Flow, a flow-matching generative framework for physical field modeling on both regular and irregular geometries. Its key design is to use a truncated POD (Proper Orthogonal Decomposition) basis — computed once via SVD of training snapshots and kept fixed thereafter — as the output representation of a POD Neural Operator (PODNO) backbone, so that the network learns the generative dynamics only in the low-dimensional POD coefficient space and is naturally decoupled from any specific mesh discretization. The model is trained with a clean-endpoint (x-prediction) objective, and a DPS-style observation-guidance term is added at sampling time, allowing a single trained model to perform forward prediction, sparse reconstruction, and inverse problems without retraining. A theoretical analysis of the truncation rank is provided, deriving a criterion G(K) that balances truncated signal energy against noise power and suggesting that energy-based truncation also suppresses a substantial fraction of the interpolation-path noise. Experiments on four benchmarks — Darcy flow and Burgers' equation (regular grids), and pipe flow and 3D geological linear elasticity (irregular meshes) — show that POD-Flow is competitive with DiffusionPDE, FM-U-Net, and FM-FNO on regular grids and clearly outperforms CoNFiLD on irregular geometries, with ablations on mode count, network depth, and prediction target, an SNR(t) analysis supporting the x-prediction design, and anonymized code released for reproducibility.

Critical strengths

S1. A simple, unified, and empirically strong framework for generative PDE solving on irregular geometries — the setting where most competing methods are weakest. POD-Flow's core move — fixing a truncated POD basis as the "output coordinate system" of a flow-matching model, so the network only learns dynamics in coefficient space — is elegant in its simplicity and directly removes the regular-grid constraint that limits DiffusionPDE/FM-FNO-style approaches. Combined with the DPS-style observation-guidance term, a single trained model handles forward prediction, sparse reconstruction, and inversion without retraining. The empirical payoff is largest exactly where it matters: on the irregular-mesh pipe-flow (918 nodes) and 3D geological elasticity (13,481 nodes, 6 coupled fields) benchmarks, POD-Flow improves markedly over CoNFiLD, with consistent gains across observation densities (Figs. 15, 22). This is consequential to my assessment because irregular-geometry capability is the principal barrier between ML-PDE generative models and real engineering use (digital twins, data assimilation, inversion); the paper's strongest claim to impact rests on this, and the evidence for it is comparatively solid. S2. Unusually complete experimental disclosure and reproducibility for this submission class. The appendices document dataset construction in full (including a purpose-built 3D geological elasticity benchmark with layer-wise random material fields), leakage-safe temporal splitting before window extraction, per-model hyperparameters and learning rates, per-variable error-vs-observation-density curves, ablations over mode count K, depth L ∈ {1,2,4,6}, and prediction target (x- vs v-prediction), plus an SNR(t) analysis that mechanistically corroborates the x-prediction design choice. Anonymous code is released. This is consequential because it makes the headline claims independently verifiable, raises confidence that the reported margins are not artifacts of undisclosed tuning, and gives the community a reusable benchmark — factors that materially push me toward acceptance despite the weaknesses below.

Critical weaknesses

W1. The inverse-problem value proposition is validated only under idealized conditions: noise-free observations and purely point-estimate evaluation. Appendix A.5 states that all reconstruction/inversion experiments use noise-free observations. Yet the entire motivation for a guided generative inverse solver is robustness to noisy, incomplete real-world measurements, and DPS-style guidance is known to be sensitive to observation noise and to the guidance scale ζ (which is never studied). Compounding this, although the framework is generative and inverse problems are ill-posed with inherently multimodal posteriors, every reported metric is a single-sample relative L2 — no posterior diversity, calibration, or coverage is evaluated. This is consequential because these two properties (noise robustness, posterior quality) are precisely what would justify a generative framework over a deterministic operator baseline; as written, the paper's central claim for the inverse setting is asserted rather than demonstrated. This is my primary reason for not scoring higher. W2. The theory is disconnected from practice, and the irregular-mesh comparison rests on a single, possibly under-tuned baseline. The Section 4.4 truncation analysis derives an optimal rank K* = argmin G(K), but no experiment ever computes or uses K* — all datasets use a 99.9%-energy heuristic instead — and the analysis assumes isotropic noise along a linear interpolation path, which does not cover the actual guided (nonlinear DPS) sampling dynamics. The theory therefore reads as post-hoc rationalization rather than a design tool. Separately, on irregular geometries the only baseline is CoNFiLD, and its U-Net configuration for the geological task (base channel width C=1, Table 3) is implausibly small; no deterministic mesh-native operator (GNO/GNOT/Transolver-class) is compared on the forward task. This is consequential because it leaves open two possibilities that would each weaken the paper: that the theory adds no actionable guidance, and that part of the reported margin reflects baseline weakness rather than method quality. Both are addressable within a revision cycle, which keeps this at Weak Accept rather than lower.

Questions for authors

Clarification requests (answers could materially change my rating): Q1. Can you provide inverse-task results with noisy observations — e.g., SNR ∈ {40, 30, 20} dB on at least Darcy (regular) and pipe flow (irregular) — including sensitivity to the guidance scale ζ? Test-time noise injection without retraining would suffice. If performance degrades gracefully and ζ admits a simple selection rule, this directly resolves W1's first half and would move me toward a higher score. Q2. Please justify the CoNFiLD U-Net configuration on geological elasticity (C=1, Table 3). Was this dictated by memory limits? Can you report a larger-capacity CoNFiLD, or a comparison matched in parameter count / wall-clock / memory, so the irregular-mesh margins can be attributed to the method rather than baseline capacity? Q3. What K* does your criterion G(K) actually predict on each dataset, and how does it differ from the 99.9%-energy choice (e.g., Darcy K=10033, pipe K=916)? Does a model trained at K* perform comparably? If the criterion is not meant to be used quantitatively, please say so explicitly and restate the theory's role as qualitative explanation. Q4. How does the fixed POD basis behave under mismatch — e.g., basis built from coarse-mesh snapshots applied at finer resolution, or test samples with significant energy in truncated modes? A single such experiment would substantiate the claimed generality boundary of the "framework." Suggestions for future work (not required for acceptance, not part of my rating): Q5. Posterior-level evaluation of the generative inverse solver (sample diversity, calibration/coverage, CRPS) versus CoNFiLD; learned or adaptive bases versus fixed POD; and autoregressive temporal modeling in place of time-as-channel. These would strengthen the probabilistic story but exceed what I consider necessary for this submission.

Reviewer_hkU9OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes POD-Flow, which combines the existing PODNO with flow matching, endpoint prediction, and observation guidance. It aims to handle both forward and inverse PDE problems on regular and irregular meshes. The POD basis is constructed from training data to capture global field information. Instead of directly predicting the flow velocity, the model predicts the clean physical field. Observation guidance is then used during inference to handle different tasks without retraining.

The authors evaluate the method on four PDE benchmarks, including Darcy flow, Burgers' equation, Navier–Stokes, and 3D linear elasticity. The results show competitive reconstruction accuracy on regular grids and better performance on several irregular-mesh tasks. The method also uses fewer sampling steps than the diffusion-based baselines.

However, the POD basis is fixed and depends on the training data and discretization. The paper does not demonstrate generalization to new meshes or geometries. All experiments also use noise-free observations. Long-term prediction is not evaluated, as the model processes temporal data using fixed-length windows..

Critical strengths

POD-Flow achieves better results on some reconstruction metrics, especially on irregular meshes. It also shows improved sampling efficiency, requiring fewer sampling steps than the diffusion-based baselines.

Critical weaknesses

I'm not fully convinced that POD is the right choice for flow matching. The paper shows that POD preserves the most signal energy for a given rank and analyzes its effect on Gaussian noise. But this doesn't necessarily mean PODNO is a better generative backbone. It's still unclear how this translates into better generation accuracy, inverse reconstruction, or sampling efficiency.

I also think preserving more energy doesn't necessarily mean preserving the most important physical details. The authors actually show that aggressive POD truncation can lose useful information. FNO and U-Net even achieve higher endpoint prediction SNR in the later stages of generation. So the advantage of POD seems to be mostly in the intermediate denoising stages, rather than throughout the whole process.

Another concern is the fixed POD basis. It's computed from the training data and stays unchanged during training and inference. The paper doesn't show whether the model can handle different meshes or geometries. To me, this looks more like a generative model designed for a fixed discretization, rather than a neural operator that can generalize across different meshes.

Questions for authors
  1. What actually makes the method better? Is it mainly POD, endpoint prediction, or observation guidance? I'd like to see how much each part contributes.
  2. Can the model work on different meshes or geometries? The experiments only use fixed meshes. Also, in the Pipe Flow experiment, almost all POD modes are kept. Does POD still offer much benefit in this case?
Reviewer_r7vWOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

POD-Flow combines an existing Proper Orthogonal Decomposition Neural Operator (PODNO) with clean-endpoint flow matching and observation-guided sampling for physical-field generation. A fixed POD basis constructed from training data captures global spatial structure, while a parallel pointwise branch retains access to information outside that basis. Learning the joint distribution of parameters and solutions allows one trained model to support different observation configurations. Contributions include this integration, an energy-based explanation of POD truncation, and evaluation on Darcy flow, Burgers’ equation, pipe flow, and geological elasticity. Results show competitive regular-grid reconstruction and substantial pipe-velocity improvements with fewer sampling steps. Validation covers fixed discretizations and noise-free observations.

Critical strengths
  • Meaningful reconstruction improvements on irregular meshes. The strongest evidence is the pipe-flow experiment: Table 1 reports consistently lower velocity errors than CoNFiLD across observation counts. At 368 observation nodes, the u-velocity error decreases from 0.4221 to 0.2090, using 100 rather than 1,000 sampling steps. This supports the practical value of the approach on the evaluated mesh.

  • A flexible inference framework with broad task coverage. Introducing observations during sampling allows the same trained joint model to accommodate different observation sets. The four benchmarks span regular grids and irregular meshes, including three-dimensional elasticity. Constructing POD bases exclusively from training data and splitting pipe trajectories before extracting windows are appropriate safeguards against data leakage.

Critical weaknesses
  • The POD-specific advantage and practical efficiency are not sufficiently isolated. The backbone comparisons do not establish performance under matched prediction targets, training budgets, and sampling budgets. The endpoint-versus-velocity ablation is limited to PODNO, with the quoted reconstruction improvements illustrated by individual examples in Figure 23. Moreover, Equation 25 analyzes a linear projection proxy, rather than the full nonlinear network with its compensation branch. The pipe model retains 916 modes for 918 nodes, making substantial noise rejection through spatial truncation an unlikely explanation for its strongest results. Finally, fewer sampling steps and trainable parameters do not establish lower latency or memory usage without accounting for guidance gradients, fixed-basis storage, and POD preprocessing.

  • The probabilistic inverse-solving claims lack corresponding validation. Reconstruction errors and denoising SNR establish pointwise predictive performance, but do not establish conditional uncertainty quality or posterior fidelity. Equation 13 supplies observation guidance without demonstrating that the resulting samples follow the desired conditional distribution. This matters especially for ill-posed material inversion: the geological material distributions are relatively concentrated, yet no prior-mean baseline establishes how much information is recovered from displacement observations. Young’s modulus reconstruction also remains worse than CoNFiLD at every observation count in Table 1.

Questions for authors
  • Do the advantages persist when PODNO, FNO, and U-Net use the same endpoint-prediction objective and comparable training and inference budgets? Please provide accuracy-versus-runtime comparisons, including guidance computation and fixed-basis memory.

  • How much does POD truncation contribute on pipe flow, where 916 of 918 spatial modes are retained? Would full-basis and reduced-mode ablations support the proposed explanation?

  • Are reported reconstructions individual samples or averages over samples? What evidence supports meaningful conditional uncertainty, and how does geological parameter recovery compare with a layer-wise prior-mean estimator? Please also clarify whether the randomly sampled boundary loads are known during inference.

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

Related papers

Discussion (0)

Sign in to comment.