acceptodds
Withdrawn from ICLR 2027

GeoPI-LNODE: Geometry Conditioned Physics Informed Latent Neural ODEs for Long-Horizon Prediction

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

We propose a Geometry-conditioned Physics-Informed Latent Neural Ordinary Differential Equations (GeoPI-LNODE), a continuous-time surrogate modeling framework for long-term prediction of complex unsteady systems governed by known partial differential equations (PDEs). High-fidelity numerical simulations are essential for engineering analysis and design, but they often incur prohibitive computational costs, motivating the development of learning-based surrogate models to enable efficient prediction. However, accurate long-horizon prediction using surrogate models remains challenging due to accumulated rollout errors, physical inconsistencies, and difficulties in handling complex geometries. GeoPI-LNODE combines parameter- and boundary-conditioned latent dynamics, geometry-conditioned field reconstruction, and physics-informed regularization to model complex spatiotemporal dynamics while preserving physical consistency over long-horizon rollouts. A latent neural ODE evolves a low-dimensional state, while a coordinate-based decoder reconstructs physical fields and enables physics-informed regularization by using residuals from the governing equations to regularize the latent dynamics in physical space. We evaluate GeoPI-LNODE on coupled fluid–scalar transport, nonlinear hyperelastic dynamics, and three-dimensional aneurysm flow—challenging regimes involving tightly coupled processes, strong nonlinearities, and complex three-dimensional geometries, where surrogate models often struggle to maintain accurate and stable long-term predictions. Within each fixed computational domain, GeoPI-LNODE improves long-term prediction accuracy over representative surrogate baselines and approximately halves the flow-rate imbalance in the three-dimensional flow problem relative to the physics-informed variant without geometry conditioning. These results demonstrate the potential of physics-informed latent dynamics for accurate long-horizon surrogates.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Physics-Informed Machine Learning, Neural ODEs, Latent Dynamics, Surrogate Modeling, Long-Horizon Prediction

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_S69dOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes GeoPI-LNODE, which combines latent neural ODEs, geometry-conditioned reconstruction, and physics-informed regularization for long-horizon prediction of PDE-governed systems. The method evolves a low-dimensional latent state and reconstructs physical fields through a coordinate-based decoder.

The authors evaluate the method on three problems: fluid and scalar transport, nonlinear hyperelastic dynamics, and three-dimensional aneurysm flow. The results show improved prediction accuracy over several baselines, as well as better flow-rate balance in the aneurysm experiment.

Critical strengths
  1. Combining latent dynamics with geometry-conditioned reconstruction and physics-informed regularization is a reasonable approach for long-horizon surrogate modeling. The method also supports field reconstruction on unstructured spatial discretizations.

  2. The experiments cover different physical systems, including a complex three-dimensional aneurysm geometry. I also appreciate that the authors evaluate flow-rate imbalance rather than relying only on pointwise prediction errors.

Critical weaknesses

1. Concerns about baseline fairness and experimental reliability.

I am concerned about the baseline results in Table 2. FNO and DeepONet show extremely large errors or even diverge in some experiments, but the paper provides little explanation. Since these experiments mainly involve temporal extrapolation within fixed geometries, I would like to see stronger evidence that the baselines were properly tuned and trained.

The definition of Baseline LNODE is also confusing. Section 4.1 states that only the physics-based losses are removed, but Table 3 seems to indicate that geometry conditioning is also removed. This makes the reported performance differences difficult to interpret.

In addition, all datasets are generated by the authors, and the three-dimensional experiment lacks comparisons with some relevant methods, such as GINO or Transolver. Evaluation on a public benchmark would make the results more convincing.

2. Concerns about physical consistency, statistical results, and reproducibility.

Some RMSE means and standard deviations in Tables F.2 and F.3 appear mathematically inconsistent with the reported numbers of test cases.

The physics-informed loss also uses stopgrad to ignore spatial derivatives of geometry features. Therefore, the calculated residuals are not the full PDE residuals of the reconstructed fields. It remains unclear whether the method improves satisfaction of the original governing equations.

Furthermore, the enforcement of boundary conditions and the treatment of derivative scaling after normalization are not sufficiently explained. Important numerical simulation details, including solver settings and convergence studies, are also missing. These issues make it difficult to assess the reliability and reproducibility of the results.

Questions for authors
  1. What is the exact origin of Baseline LNODE? Does it remove both geometry conditioning and physics-informed losses? How does it differ from "w/o physics-informed"? Also, how is the "w/o latent dynamics" ablation implemented?

  2. Why do FNO and DeepONet show extremely large errors or diverge in some experiments? Were these baselines sufficiently tuned and trained to convergence? Could the authors provide training curves and consider comparisons with GINO, Transolver, or a public benchmark?

  3. Some RMSE statistics in Tables F.2 and F.3 appear mathematically inconsistent. For example, Table F.2 reports 0.00195 ± 0.0173 for GeoPI-LNODE, which violates the standard deviation bound for 24 nonnegative per-case RMSE values. Could the authors clarify the calculations and provide per-case errors or evaluation scripts?

  4. Since stopgrad ignores the spatial derivatives of geometry features, does the method still reduce the actual PDE residuals when full derivatives are considered? How are boundary conditions enforced and derivative scaling factors handled after normalization? Also, could the authors provide more details on the numerical solvers, convergence studies, and actual inference speedup?

Minor comment: Two author names appear to be incorrect in the references. "Zhongang Zhang" should be "Zhongqiang Zhang" in Lu et al. (2021), and "Mattia Salvador" should be "Matteo Salvador" in Regazzoni et al. (2024).

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

This paper presents GeoPI-LNODE, combining latent neural ODE dynamics, a geometry-conditioned spatial decoder, and physics-based regularization for long-horizon PDE prediction. Experiments cover transport, hyperelasticity, and aneurysmal flow.

Critical strengths
  1. The framework offers a compact approach to predicting spatial fields over long time horizons.
  2. Experiments cover three distinct physical systems and include component ablations.
  3. The aneurysm results show improved flow balance, providing evidence beyond prediction error alone.
Critical weaknesses

1.The physical interpretation of the residual remains limited. As explicitly acknowledged, spatial differentiation freezes the geometry features and therefore omits part of the chain rule for the predicted field. Consequently, the surrogate residual need not control the actual PDE residual and can depend on how the decoder represents the same field. The paper needs a clearer justification of why this regularizer promotes physical consistency.

2.The baseline definition is inconsistent. Section 4.1 describes the baseline LNODE as using the same architecture while omitting physics regularization, yet its results differ substantially from the “without physics” ablation. Table 3 appears to associate the baseline with removing both physics and geometry features. This ambiguity affects attribution of the reported improvements.

3.The demonstrated scope is narrower than the framing suggests. Each system uses a fixed geometry and a common initial state. The experiments therefore establish parameter- and forcing-dependent prediction within fixed domains, rather than generalization across geometries or arbitrary initial conditions. The contribution relative to existing latent dynamics models should be positioned accordingly.

Questions for authors
  1. Under what conditions does the frozen-feature residual approximate or control the true PDE residual?
  2. What precisely distinguishes the baseline LNODE from the “without physics” ablation?
  3. How should the geometry-conditioning contribution be interpreted given the fixed-geometry evaluations?
Reviewer_6G1hOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies long-horizon prediction of parameterized physical systems within fixed domains. A latent neural ODE is conditioned on physical parameters and time-dependent inputs, while a coordinate decoder additionally receives geometry-related features. Training combines field reconstruction with a frozen-feature physics surrogate: spatial derivatives through the geometry features are deliberately omitted. Three separately trained models address flow-transport, hyperelastic dynamics, and three-dimensional aneurysm flow, with training/test endpoints of 20/40, 10/40, and 10/100 seconds. The paper reports improved extrapolation accuracy and reduced aneurysm inlet-outlet flow imbalance when geometry conditioning is added to physics regularization.

Critical strengths
  1. The flow-balance diagnostic adds information beyond pointwise error. Table 3 reports that adding geometry conditioning to the physics-regularized model reduces mean flow imbalance from 22.9% to 12.1%, a 47.2% relative reduction, while velocity-magnitude RMSE decreases from 0.4033 to 0.4026, about 0.17%. If confirmed, this result supports evaluating integral flow balance alongside prediction error.

  2. The evaluation covers complementary physical settings and explicit temporal extrapolation. The experiments span coupled transport, nonlinear structural dynamics, and an unstructured three-dimensional flow domain, with test endpoints two, four, and ten times the training endpoints (Section 4.2). The authors clearly limit the results to fixed geometries (Section 3.2). The application study is potentially informative within that scope.

The reported difference between improvements in velocity-magnitude accuracy and inlet-outlet balance is useful for evaluating physical surrogates, subject to verification of the results in Table 3.

Critical weaknesses

The architecture and frozen-feature convention are explicit, but the baseline description and quantitative reporting require reconciliation. The issues below may be addressable through corrected reporting and focused validation, but they currently limit confidence in the experimental conclusions.

  1. The reported error summaries are inconsistent under the stated definitions. Table F.2 reports GeoPI-LNODE beam vxv_x RMSE as 0.00195±0.01730.00195\pm0.0173 across the 24 test cases specified in Appendix E.2. For 24 nonnegative case-wise RMSEs with mean rˉ\bar r, the sample standard deviation satisfies s≤24 rˉ≈0.00955s\leq\sqrt{24}\,\bar r\approx0.00955; the population-SD bound is smaller. The reported SD of 0.0173 exceeds this limit even after allowing for rounding. Separately, Table 2 gives aneurysm Baseline LNODE mean velocity-magnitude RMSE as 4057, whereas the component means in Table F.3 sum to 1155+1138+295.6=2588.61155+1138+295.6=2588.6. With common cases, units, spatial/time evaluation points, and weights, the mean magnitude RMSE cannot exceed the sum of the mean component RMSEs. This discrepancy also exceeds rounding tolerance. These contradictions affect both the proposed model's reported variability and a prominent baseline comparison. The tables need correction or an explicit explanation of any differences in normalization or aggregation before they can support reliable performance comparisons.

  2. The component attribution and interpretation of physical consistency need clarification. Section 4.1 describes Baseline LNODE as using the same architecture and optimization as GeoPI-LNODE with physics losses removed. However, Table 3's no-physics/no-geometry row matches the baseline's aneurysm error to the reported precision, while retaining geometry without physics gives 0.4141. This suggests that the baseline comparison removes both components, which the description should state explicitly. Table 3 does provide all four combinations and supports comparisons within that design. However, the claim of improved local physical consistency in Section 5 needs separate support: Section 3.2 explicitly acknowledges that the training residual omits geometry-feature derivatives, and the additional physical diagnostic in Table 3 measures inlet-outlet imbalance. These results support the narrower flow-balance claim but do not establish improved local continuity of the reconstructed field.

The fixed-domain application study and flow-balance diagnostic have potential value. However, the incompatible error summaries, inconsistent baseline description, and limited support for the local physical-consistency claim currently weaken the empirical basis for acceptance. I would reconsider if corrected results preserve the principal findings and the component comparisons and physical claims are clarified.

Questions for authors
  1. Can you reconcile the two numerical contradictions above and provide the per-case errors underlying the affected entries, including units, evaluation points, aggregation axes, and SD convention? Please check Tables 2 and F.1 to F.3 for the same issues and explain whether the main comparisons change.

  2. Does Baseline LNODE omit both geometry features and physics losses, or only physics losses? Please specify the inputs and loss terms for each LNODE variant so that the comparisons in Tables 2 and 3 can be interpreted consistently.

  3. Does the claim of improved local physical consistency refer to the surrogate training residual or to continuity of the reconstructed field? If the latter, can you provide a physical-space continuity diagnostic for the existing aneurysm models that accounts for variation of the geometry feature, or an appropriate mesh-based diagnostic? Please also specify how boundary fluxes are integrated and whether the inlet normalization uses predicted or reference/prescribed flow. This is a request to validate the existing claim; otherwise, narrowing it to the measured flow-balance improvement would clarify the contribution.

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

Related papers

Discussion (0)

Sign in to comment.