acceptodds
Under review as a conference paper at ICLR 2027

Physics-Informed Neural PDE Solvers via Spatio-Temporal MeanFlow

Abstract

Deep learning paradigms, such as Physics-Informed Neural Networks (PINNs) and neural operators, have significantly advanced the solving of partial differential equations (PDEs). However, they often struggle to capture the continuous integral nature of physical systems, relying either on pointwise residuals that ignore the integral perspective or on pre-discretized temporal grids. Drawing inspiration from MeanFlow, a continuous-time integrator recently developed to efficiently solve generative ODEs, we introduce Spatio-Temporal MeanFlow, which functions as a novel PDE solver learning the finite-interval evolution of physical states. By substituting the generative velocity field with the physical PDE operator, we transform multi-step numerical integration into an efficient prediction with a controllable integration length. Crucially, we extend the original MeanFlow loss from the temporal to spatio-temporal domain, coupling time evolution with spatial consistency. This yields a unified framework naturally accommodating both time-dependent and stationary PDEs. Comprehensive experiments on canonical benchmarks demonstrate that our approach achieves superior accuracy over representative baselines. Furthermore, the proposed integral constraint enables excellent zero-shot generalization to out-of-distribution initial conditions and varying spatial resolutions.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
PDE solving, MeanFlow

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

This paper proposes SPATIO-TEMPORAL MEANFLOW, a physics-informed framework for learning PDE solution operators through finite-interval state evolution. It extends MeanFlow to the spatial domain and combines supervised learning with decoupled temporal and spatial constraints, accommodating both time-dependent and steady-state PDEs. The paper reports improved prediction accuracy on PDE benchmarks, alongside ablations and evaluations of flexible inference steps, spatial-resolution generalization and OOD initial conditions.

Critical strengths
  • An interesting extension of MeanFlow: The paper adapting temporal MeanFlow to PDE forecasting, and also introduces spatial regularization. This broadens the modeling perspective.
Critical weaknesses
  • The central motivation is not adequately justified: The paper argues that existing paradigms rely on pre-discretized temporal grids. I think this characterization misleading. Neural operators such as FNO claims they are designed to learn mappings between infinite-dimensional function spaces, which can directly represent evolution over an entire finite time interval. Using discretized training data does not imply an inability to represent the result of continuous evolution. The FNO and PINO baselines in this paper directly predict complete trajectories without autoregressive feedback, as stated in Appendix F.3. The claim in the abstract that these methods "rely on pre-discretized temporal grids" and "struggle to capture the continuous integral nature of physical systems" is therefore not adequately supported. The paper needs to explain why learning average rates of change over finite time intervals is preferable to directly learning the solution operator. Instead, the current motivation justifies the proposed approach by attributing limitations to the baselines that do not apply to their implementations.

  • The second-order derivation incorrectly applies the chain rule: Even in the one-dimensional case, Eq. (23) omits the terms muu(uξ)2m_{uu}(u_\xi)^2 and 2muξuξ2m_{u\xi}u_\xi, while incorrectly doubling the explicit mξξm_{\xi\xi} term. The resulting second-order identity in Eqs. (10) is not valid in general under the stated assumptions. The identity is explicitly used to construct the Navier-Stokes residuals, undermining their claimed physical consistency.

Questions for authors
  • Impact of the derivation error: Does the second-order derivation error affect the implemented training losses, particularly the Navier–Stokes residuals? Please clarify which expressions are actually used in training and whether correcting the error changes the reported results. If these expressions are not used, how does the implementation relate to the presented formulation?

  • Loss weighting: Does the proposed decoupling mitigate gradient conflicts? The paper reports common weights of λt=λs=0.01\lambda_t=\lambda_s=0.01 for the main results, how well does this choice of loss weights generalize across different PDE settings? To substantiate the claimed optimization benefits, please provide detailed metrics characterizing the training dynamics, such as how the weighted gradient norms of individual loss terms and the pairwise cosine similarities between their gradients evolve throughout training.

  • Consistency across integration intervals: Does the learned operator produce the same final state when an interval is predicted directly versus divided into smaller subintervals? Do the proposed losses enforce or otherwise control this composition consistency? Quantifying the discrepancy would help establish whether the model learns a consistent evolution operator.

Reviewer_YZ8kOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper proposes to adapt MeanFlow, a one-step generative modeling method, to solving PDEs. The authors extend the MeanFlow identity from time to space-time and derive physics-informed constraints. These are splitted into a temporal loss and a spatial loss to avoid ill-conditioned optimization.

Trained with supervised data plus these losses, the Meanflow-based method gets the lowest error among the methods tested on Burgers, Navier–Stokes, Darcy and Poisson. It is also achieving the best performance, beating FNO, PINO, PINNs and CFO. It also generalizes zero-shot to higher resolutions and out-of-distribution initial conditions on Burgers.

The authors also conduct ablations to check what each MeanFlow loss adds, how sensitive the method is to the hyperparameter in the loss and how inference and generalization behave.

Critical strengths

(1) The authors turn MeanFlow’s finite-interval average from generative modeling into a physics-informed PDE solver.

(2) Splitting the loss into temporal and spatial parts is supported by an error bound and a decoupling bound. The authors also state their assumptions and limitations openly.

(3) The proposed method has the lowest error among the tested baselines on all four benchmarks. In addition, inference time is little, showing the efficiency of the proposed method.

Critical weaknesses

(1) Errors in writing: Appendix F.4 points to Tab. 1, but the results are in Table 5. A duplicated sentence in Appendix F.5. “Three representative paradigms”, followed by a list of two.

(2) The scope of the project is a bit limited: There are some existing works in the literature that maps (u(τ),τ,t)(u(\tau), \tau,t) to u(t)u(t). Flow Map Matching (Boffi, Albergo, Vanden-Eijnden, 2024) for example, is not properly cited or compared. MeanFlow is closer to these types of works. The authors have also failed to cite these works in related works: flow-map or consistency-model, DiffusionPDE (even though its data protocol is used), No transformer operators (Transolver, OFormer), etc. CViT and Geo-FNO are only mentioned in the appendix.

(3) The experiment sessions show strange and confound configurations of underlying architectures. On Navier–Stokes, Darcy and Poisson the method uses a Fourier U-Net, while the baseline is a standard FNO. FNO+MeanFlow loss and Fourier U-Net+PINO-style residual are the missing.

(4) While the authors claim that decoupiling is a contribution with its own theorem. The individual decoupling effect is not shown in the ablation. The coupled loss is never tested against the split losses.

(5) The baselines are too easy. For Navier–Stokes, T=1T=1 at ν=1e−3ν=1e-3 is a short, mild horizon. So the gains sit at the 10−310^-3 level. Darcy and Poisson use a fixed, homogeneous zero boundary. Burgers has ν=0.01ν=0.01, so it’s smooth with no real shocks. The authors should consider switching to much harder benchmarks instead of self-generated data, if they want to claim the usefulness of the method.

(6) The authors give many training budgets to turn their method's own learning rates and epochs (100–300) are tuned per dataset. The baselines simply “follow official implementations”. The tuning budgets are not comparable and even across baselines and the claimed method.

Questions for authors

(1) For the experiments that are averaged across 3 seeds, why do several results report ±0.00\pm0.00 ?

(2) Have the authors consider error-over-time curve, energy spectrum or vorticity statistics for Navier–Stokes?

(3) Have the authors used other metrics like RMSE or MAE than just the relative L2 error?

(4) Do the MeanFlow losses use stop-gradient on the target side, as the original MeanFlow does? If not, how is training kept stable?

(5) How is KK computed in the practice? Is KK ever close to singular during training? Do you check or regularize for that?

(6) Can you report “same backbone + data loss only” next to FNO, to separate the effect of the architecture from the effect of the losses?

Reviewer_S73HOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces Spatio-Temporal MeanFlow, a physics-informed neural PDE solver that learns the evolution of physical systems over continuous time intervals. The method combines spatial and temporal physics constraints to improve prediction.

Critical strengths
  1. The paper presents a approach by extending MeanFlow to solve PDEs using spatial and temporal physics-informed constraints.
  2. The method can handle both time-dependent and steady-state PDEs.
  3. The proposed method achieves better accuracy than several existing PINN and neural operator methods.
  4. The method shows good generalization to different spatial resolutions and unseen initial conditions.
  5. The framework allows flexible integration steps during inference without requiring retraining.
Critical weaknesses
  1. The experiments focus on 1D and 2D PDEs. How well the method performs on more complex high-dimensional problems.
  2. Although some experiments on irregular domains are provided, the performance on complex geometries, unstructured meshes, and different boundary conditions is not sufficiently explored.
  3. The proposed method involves additional spatial and temporal derivative calculations, which may increase computational costs during training. A more detailed comparison of training time, memory requirements, and computational efficiency would be helpful.
  4. The theoretical analysis assumes smooth PDE solutions. It is unclear how well the method performs for problems involving shocks, discontinuities, or sharp changes in the solution.
  5. The paper provides limited analysis of long-term prediction stability and error accumulation over longer time intervals.
Questions for authors
  1. How does the proposed method perform on high-dimensional PDE problems? Can the authors provide additional experiments to demonstrate its scalability beyond the primarily 1D and 2D benchmarks?
  2. Although the authors provide some experiments on irregular domains, how well does the method generalize to more complex geometries, unstructured meshes, and different boundary conditions?
  3. The proposed method requires additional spatial and temporal derivative computations during training. Can the authors provide a more detailed comparison of training time, GPU memory usage, and computational cost with existing methods, particularly as the problem dimension and spatial resolution increase?
  4. How does the proposed method perform on PDEs with shocks, discontinuities, or sharp gradients, given that the theoretical analysis assumes sufficiently smooth solutions?
  5. Can the authors provide additional experiments to evaluate the long-term stability of the method and investigate error accumulation over longer prediction horizons?
  6. Can the authors compare the same backbone trained with standard PDE residual regularization instead of MeanFlow losses? This would help clarify whether the improvements come from the proposed MeanFlow formulation rather than the backbone architecture or physics-based regularization in general.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.