acceptodds
Under review as a conference paper at ICLR 2027

Generalized Neural Operator for Parametric and Boundary-Value Problems

Abstract

Many physical systems are governed by Partial Differential Equations (PDEs), where dynamics are dictated by underlying physical parameters and boundary conditions. Modeling real-world phenomena requires a unified solver capable of generalizing across an entire family of PDE instances, yielding instant solutions for various parameters or boundary constraints without necessitating architectural redesign or retraining. However, existing approaches face a fundamental trilemma, failing to simultaneously achieve computational efficiency, mathematical well-posedness, and generalization. Traditional numerical solvers guarantee well-posedness but are computationally expensive and strictly instance-specific. Physics-Informed Neural Networks (PINNs) offer fast inference and physical fidelity, yet they remain instance-specific and require costly retraining for every new condition. Purely data-driven Neural Operators generalize across instances but treat physics implicitly. By attempting to infer parameters and boundaries from historical frames, they lead to ill-posed formulations. Recent large-scale foundational operators inherit this implicitness while scaling to billions of parameters, largely negating their computational advantage over optimized numerical baselines. To resolve this trilemma, we propose a Generalized Neural Operator that restores mathematical rigor through explicit physical conditioning. By formalizing the classical conditions for well-posedness within neural operators, our framework demonstrates the theoretical benefits of explicitly conditioning on PDE parameters and boundary conditions. To implement this synthesis without compromising computational speed, we introduce three novel architectural components: a parameter-gated mixture of kernels for efficient parameter generalization, a generalized boundary transfer operator that projects arbitrary boundary constraints into a unified latent Dirichlet representation, and a specialized training objective to ensure stability. Extensive experiments demonstrate that our theoretically grounded approach achieves superior generalization across heterogeneous physical regimes while maintaining inference efficiency comparable to conventional numerical baselines.

Primary area:
learning on time series and dynamical systems
Keywords:
physics informed neural network

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_gd4dOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

The authors present a method for constructing and training neural operators (NO) to improve generalisability across parameters and boundary conditions. This is done by using a PDE parameter-dependent kernel in the NO, embeddings of BCs as DIrichlet BCs, and with a minimax-type loss. Several PDEs are then tested and cost accuracy is compared with standard numerical schemes.

Critical strengths
  • The paper addresses an important problem in surrogate modelling for PDEs, namely generalisability and robustness to changes in PDE parameter and boundary conditions. This is an under-developed area of research in this field. The presented method seems to propose a simple yet sensible way of improving baselines meaningfully. Furthermore, they perform good comparisons with standard numerical methods.
Critical weaknesses
  • The paper lacks clear mathematical exposition of certain aspects of the method which lead to ambiguity and may make reproducibility an issue. Maps which could be described more precisely are explained with words and references which make the method obscure.
Questions for authors
  • Could you use something other than angle brackets for concatenation? This is quite confusing.

  • How is the boundary condition map implemented precisely?

  • And could the same results be achieved simply by using a different network for each PDE parameter "bin"?

4 comments

We thank you for your insightful comments on improving our paper. We hope the following could address your concerns.

Q1 Concatenation Notation

We will denote concatenation by concat(⋅)concat(\cdot). The first term of Equation 2 then reads Wconcat(θ~,h(l),v(l)(x))W concat(\tilde{\theta}, h^{(l)}, v^{(l)}(x)). We thank the reviewer for suggesting this change and it will improve the readability of the equations.


Q2 Boundary Condition Map

The boundary encoder patchifies the boundary data along the domain edges, analogous to the patch embedding in ViT, and embeds each boundary patch as a boundary token. The boundary type (periodic, Dirichlet, or Neumann) is mapped to a learned type embedding. These are concatenated with the current latent state tokens, and a self-attention layer updates the boundary tokens using both the type embedding and the interior latent state. The updated boundary tokens form the latent Dirichlet representation. They are prepended to the latent state tokens together with a PDE parameter token produced by an MLP, and the sequence is processed by the kernel selected by the parameter gate. The boundary and parameter tokens are then removed, so only the latent state tokens pass to the next layer.


Q3 Separate Network per Parameter Bin

A separate network per parameter bin differs from our mixture of kernels in three ways. First, each layer has its own gate, so the number of kernel combinations grows exponentially (KLK^L, where KK is number of kernel and LL is the number of layers) with depth while the number of kernels grows only linearly (K∗LK*L). Covering the same combinations with separate networks would require significantly more networks, which could be infeasible to train and store.

Second, all components other than the attention kernels, including the encoder, decoder, MLPs, boundary transfer operator, and parameter embedding, are shared across all parameter values. These shared components learn the structure common to all parameter regimes so every parameter region benefits from the data of all others. A separate network per bin only sees the data in its own bin which could be insufficient to learn good representations and is also computationally expensive.

Third, the partition of the parameter space is learned from data and composed across layers. The resulting partition is not known in advance. Having separate networks would require choosing bins by hand and injecting human priors about where the physical regimes change, which are generally unknown. This would overly rely on human designers and could also limit the model’s expressiveness.

To definitely answer this question, we report a comparison with separate networks trained on 10 bins of the incompressible Navier–Stokes parameter range. The results show that the single pretrained model has much better performance than the single networks.

SubsetTrain ReTest RenMSE
0100–500100–5508.36×10−18.36 \times 10^{-1}
1600–1000575–10507.57×10−17.57 \times 10^{-1}
21100–15001075–15507.96×10−17.96 \times 10^{-1}
31600–20001575–20508.36×10−18.36 \times 10^{-1}
42100–25002075–25508.08×10−18.08 \times 10^{-1}
52600–30002575–30508.42×10−18.42 \times 10^{-1}
63100–35003075–35508.27×10−18.27 \times 10^{-1}
73600–40003575–40508.90×10−18.90 \times 10^{-1}
84100–45004075–45508.50×10−18.50 \times 10^{-1}
94600–50004575–50008.26×10−18.26 \times 10^{-1}
Full Model (Ours)100-5000100-50001.06×10−1\mathbf{1.06 \times 10^{-1}}

W1 Explanation on Function Maps

We have revised Sections 4.2 and 4.3 so that every map in the model is defined by an explicit equation, with the changes highlighted in red. The encoder splits the input field into non-overlapping patches and embeds them with an MLP, and the decoder applies an MLP and reassembles the patches into the field. The PDE parameters are embedded by an MLP, and a separate MLP produces one score for each kernel. Each base kernel is a self-attention layer with its own query, key, and value weights, so the kernel in Equation 2 is given by the attention weights. The parameter and boundary tokens are added in front of the latent state tokens, serve as the context set in Equation 2, and are removed after each layer. The boundary transfer operator embeds the boundary data along the domain edges as boundary tokens, maps the boundary type to a learned embedding, and updates the boundary tokens with a self-attention layer over the boundary and interior tokens to form the latent Dirichlet representation. The gate selects the kernel with the highest score in the forward pass, and a straight-through estimator passes gradients through the softmax probabilities in the backward pass, so the partition of the parameter space is learned during training. Appendix S gives the full pseudocode, and we will release our code upon publication.


Thank you for your constructive feedback. We have updated the manuscript to incorporate your suggestions. We hope this clears up any confusion and concerns, and if so, we would greatly appreciate your consideration in raising the score.

I thank the authors for their comprehensive response and clarifications. These help the clarity of the paper and answer my concerns.

I have increased my score accordingly.

We sincerely thank the reviewer for the positive feedback. We greatly appreciate your support and your high evaluation of our work.

Reviewer_281AOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a generalized neural operator (GNO) for learning PDE solution operators across varying physical parameters and boundary conditions. The method combines a parameter-gated mixture of kernels, a generalized boundary transfer operator for encoding different boundary conditions, and a group DRO training objective. Experiments on heat, advection, burgers, and NS equations demonstrate improved accuracy over several baselines.

Critical strengths
  1. The paper addresses an important problem of generalizing neural operators across PDE parameters and boundary conditions without retraining.
  2. The proposed parameter-gated mixture of kernels provides an interesting approach to adapting neural operators across different physical regimes while maintaining inference efficiency.
  3. The experiments cover multiple PDE families and demonstrate consistent improvements over baselines. The accuracy-matched runtime comparisons with numerical solvers are also valuable.
Critical weaknesses
  1. The role and effectiveness of individual components are not sufficiently justified. The ablation study is limited to one PDE setting and does not fully isolate the contributions of the proposed components.
  2. Comparisons with standard approaches are limited. For example, PINO is only compared in the appendix on heat and advection, without explicit parameter conditioning at inference.
  3. The theoretical contribution is limited. Proposition 4.1 mainly establishes the necessity of sufficient physical conditioning, without providing well-posedness or stability guarantees for the learned operator.
Questions for authors

Why are varying boundary condition types only evaluated on heat and advection, rather than more challenging nonlinear PDEs such as burgers and NS equations?

8 comments

We thank you for your insightful comments on improving our paper. We hope the following could address your concerns.

W1 Ablation Study and Role of Individual Components

To achieve generalization across PDE parameters and boundary conditions, we propose three core mechanisms: (1) a parameter-gated mixture of kernels, (2) a generalized boundary transfer operator, and (3) a robust optimization training curriculum. First, because PDE parameters dictate the underlying physical regimes, we utilize a parameter-gated mixture of kernels to dynamically apply different integral kernels to distinct regimes. Second, for boundary condition generalization, we draw inspiration from classical Dirichlet-to-Neumann and Neumann-to-Dirichlet maps. While these mathematical operators are traditionally restricted to specific time-independent elliptic PDEs, we generalize this concept using neural networks by learning a generalized boundary transfer operator that projects arbitrary boundary conditions into a shared latent Dirichlet representation. This allows the core neural solver to process only a single, standardized boundary condition type. Finally, neural solvers often exhibit high variance across physical regimes (e.g., performing well on some Reynolds numbers but failing on more challenging ones). To mitigate this, our robust optimization training curriculum ensures stable and consistent performance across all regimes.

In addition to the ablation study on the advection equation reported in the main paper, we present an ablation study on the heat equation below, evaluating the removal of each component one at a time. The results demonstrate that omitting any single component degrades performance by a large margin, confirming the necessity of all three contributions.

ModelnMSE
Full Model3.62×10−1\mathbf{3.62 \times 10^{-1}}
− Robust Optimization Objective5.06×10−15.06 \times 10^{-1}
− Generalized Boundary Transfer Operator1.00×1001.00 \times 10^{0}
− Mixture of Kernels9.00×10−19.00 \times 10^{-1}

W2 Comparisons with Standard Approaches

Our comparison follows the original formulation of PINO [1], in which the model input for time-dependent PDEs is the initial condition and the PDE parameters are not provided as inputs. They enter only through the PDE loss during training.

Our main comparisons cover the standard approaches to encoding physics. For PDE parameter generalization, Concat follows the foundational model [9] and concatenates the PDE parameters with the initial condition, CAPE+FNO, CAPE+Unet, and CAPE+Attn adopt the parameter-conditioning mechanism of CAPE [1], and Unisolver [2] takes the PDE parameters as explicit input. The ViT variants reflect how existing foundation models infer the physics implicitly from initial frames.

For boundary condition generalization, Unisolver takes the boundary conditions as explicit input, while the ViT variants again infer them from initial frames. We further add two concatenation strategies, Concat-Edge and Concat-Broadcast. Boundary-specific neural operators do not apply to this setting, since BOON [3] enforces each boundary type through type-specific kernel modifications and does not support single model generalization. BENO [4] is limited to time-independent elliptic PDEs.

The small number of existing methods that explicitly encode both PDE parameters and boundary conditions is not a limitation of our work. It shows that explicit conditioning on PDE parameters and boundary conditions remains an important yet understudied problem, which is precisely the gap our work addresses.


W3 Theoretical Contribution

The importance of Proposition 4.1 lies not in mathematical complexity but in its consequences for how neural operators are formulated. It is the theoretical foundation for why we champion explicit PDE parameter and boundary condition inputs. It is rooted in classical PDE theory from Hadamard well-posedness, and we formulate this requirement in the neural operator setting. This yields implications beyond the statement itself. First, the initial condition alone is insufficient to predict the future trajectory. Yet for time-dependent PDEs, neural operators are commonly formulated as a mapping from initial conditions to solution trajectories; Proposition 4.1 implies that this domain must be augmented to include PDE parameter and boundary condition space. Second, any model without these explicit inputs must infer them from the initial states, which introduces irreducible error, and for time-dependent boundary conditions this inference is impossible. These implications directly motivate our explicit conditioning on PDE parameters and boundary conditions.

W3 Theoretical Contribution (Continued)

Regarding well-posedness of our solver, existence and uniqueness are trivial, since our model maps every combination of initial condition, PDE parameters, and boundary conditions to exactly one output. Regarding stability, error bounds for traditional numerical solvers rely on the one-step map being Lipschitz, which neural solvers in general are not guaranteed to satisfy. However, on the bounded region covered by our training and test data, our model and the baselines are Lipschitz with respect to the state, since their gradients are bounded there. This gives a weak local stability guarantee. The rollout error after a given number of steps is bounded by the one-step error multiplied by a factor that depends only on the local Lipschitz constant and the number of steps.

We follow existing works to demonstrate stability empirically. All our experiments evaluate 10-step rollouts, and our model achieves the lowest rollout error across them.


Q1 Extension of Boundary Condition Variation to Other PDEs

Generating data for a single PDE under Dirichlet, Neumann, and periodic boundary conditions cannot be done by switching a setting in one solver, because each boundary type calls for a different numerical solver. The 2D viscous Burgers equation is a clear example. For periodic boundaries, a Fourier pseudo-spectral method is the natural choice, since Fourier modes satisfy periodicity automatically and give spectrally accurate derivatives. Fourier modes cannot represent walls, however, and using them on a domain with walls produces spurious oscillations at the boundary. Walls are instead handled naturally by the finite element method, where a Dirichlet condition is imposed by fixing the solution at the wall nodes and a Neumann condition enters the weak form as a boundary integral. One PDE with three boundary types therefore requires building, stabilizing, and verifying separate solvers. They must also be designed to produce data on the same grid and at comparable accuracy, so that the boundary condition is the only thing that changes across datasets.

Existing benchmarks do not provide this kind of controlled variation. In widely used neural PDE solver benchmarks [5,6,7,8], each PDE comes with one boundary setting. None of them offers a single PDE under all three of Dirichlet, Neumann, and periodic boundary conditions. Building a boundary condition study from these benchmarks would mean taking each boundary type from a different PDE. The governing equation, the parameter range, and the resolution would then all change together with the boundary condition, and any difference in accuracy would be confounded rather than attributable to the boundary condition. For this reason we generated all datasets with boundary condition variation ourselves. Each of them required careful solver analysis and verification to guarantee high-fidelity data.

Because of the tremendous effort in writing numerical solvers, large-scale datasets in which a single PDE varies jointly in its parameters and its boundary conditions call for a sustained community effort. This is a central long-term goal of our research, and we have already continued expanding our datasets in this direction. Below we provide experiment results on a new 2D viscous Burgers dataset that varies both the viscosity and the boundary conditions.

ModelnMSE
Ours2.86×10−1\mathbf{2.86 \times 10^{-1}}
Unisolver3.12×10−13.12 \times 10^{-1}
MoE-POT3.32×10−13.32 \times 10^{-1}
ViT24.34×10−14.34 \times 10^{-1}
ViT54.14×10−14.14 \times 10^{-1}
ViT103.62×10−13.62 \times 10^{-1}

Thank you for your constructive feedback. We have updated the manuscript to incorporate your suggestions. We hope this clears up any confusion and concerns, and if so, we would greatly appreciate your consideration in raising the score.

[1] Learning Neural PDE Solvers with Parameter-Guided Channel Attention, ICML 2023

[2] Unisolver: PDE-Conditional Transformers Towards Universal Neural PDE Solvers, ICML 2025

[3] Guiding continuous operator learning through Physics-based boundary constraints, ICLR 2023

[4] BENO: Boundary-embedded Neural Operators for Elliptic PDEs, ICLR 2024

[5] PDEBENCH: An Extensive Benchmark for Scientific Machine Learning, NeurIPS 2022

[6] The Well: a Large-Scale Collection of Diverse Physics Simulations for Machine Learning, NeurIPS 2024

[7] FD-Bench: A Modular and Fair Benchmark for Data-driven Fluid Simulation, KDD 2026

[8] APEBench: A Benchmark for Autoregressive Neural Emulators of PDEs, NeurIPS 2024

[9] Towards foundation models for scientific machine learning: Characterizing scaling and transfer behavior, NeurIPS 2023

Thank you for the detailed response. I agree with the theoretical motivation, although the current theorem seems rather straightforward and does not provide a substantive guarantee for the proposed neural solver.

Regarding the experiments, I believe the core challenge, and therefore the main focus of the evaluation, should be joint generalization across PDE parameters and boundary conditions, especially different boundary condition types, since parametric PDE learning has already been widely studied. However, the main paper only considers joint variation for the heat and advection equations, while most other experiments vary parameters or boundary conditions separately. While constructing datasets can be time-consuming, established numerical schemes already exist for many of the PDEs considered here, and different boundary conditions do not necessarily require entirely different solvers. I find the claim that generating such datasets requires tremendous effort and sustained community work a bit overstated, given that more realistic problems such as car aerodynamics studied in Transolver are not considered in the paper.

After reexamining the results, another major concern is the validity of the baselines and the actual predictive performance. The paper reports nMSE rather than the more commonly used relative L2 error. In Table 4, the proposed method achieves 0.362 nMSE on the heat equation, which still represents substantial error relative to the target's variation (equivalent to an RMSE approximately 60% of its standard deviation under the usual normalization). Some baselines perform very poorly, with ViT-2 reaching 26.0 nMSE, while predicting the exact target mean would yield nMSE = 1 under the standard normalization. More concerningly, Unisolver reports 0.871 nMSE on heat and 0.502 on advection. Although not directly comparable across datasets and metrics, Unisolver reports relative L2 errors on the order of 10^-2 on several benchmarks in its original paper. Such large errors here raise concerns about the baseline implementation, training, and fairness of the comparisons, particularly since Unisolver already supports explicit conditioning on PDE parameters and boundary conditions.

I agree that jointly conditioning neural operators on PDE parameters and ICBCs is a promising direction. However, given the current experimental setup, insufficient justification of the reported performance, concerns about baseline fairness, and limited reproducibility details, I am inclined toward rejection.

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

This paper proposes a Generalized Neural Operator for families of PDEs where the PDE parameters and boundary conditions vary. The authors argue that existing approaches can’t be well-posed, efficient and generalizable at the same time. They state that an operator taking only the initial condition as input can’t represent the full family of solution operators, so the PDE parameters and boundary conditions need to be explicit inputs. They contribute three architectural/training components. First, a parameter-gated mixture of kernels, where the PDE parameters select one attention kernel per sample. Second, a generalized boundary transfer operator that maps arbitrary boundary conditions to a latent Dirichlet representation. Third, a Group DRO objective to make errors more uniform across parameters. They validate the method on the heat equation, advection equation, Burgers equation and incompressible Navier-Stokes equation. Generalization is tested across parameters, across boundary conditions, and across both jointly, with additional 3D, out-of-distribution and public dataset experiments. They report lower MSE than the baselines and a 2-4x speedup over accuracy-matched numerical solvers.

Critical strengths

Strength 1: Clarity of the paper.
 The paper is well-written and easy to follow. Table 1 nicely highlights the conditions a desirable PDE solver should satisfy (well-posedness, efficiency and generalization) and where existing approaches fall short. This matters to my assessment because it makes the motivation and positioning of the work clear, and the reader can easily see what gap the paper is trying to fill.

Strength 2: Motivation for giving boundary conditions as input. 
The paper makes a convincing case for why boundary conditions should be explicit inputs to the model. Time-dependent boundary conditions are a good example: future boundary values can’t be inferred from previous time frames, so a model relying on implicit inference will fail in that setting. This is consequential because most current neural operators and foundation models don’t take boundary conditions as input, and this argument gives a solid reason for doing so.

Critical weaknesses

Weakness 1: Boundary conditions are given as input but not satisfied. The paper’s main argument is that a well-posed operator must take the boundary conditions as input. However, the model doesn’t satisfy the boundary conditions exactly, and the authors acknowledge this in Appendix O. If the boundary conditions are only used as a soft conditioning signal and the output can still violate them, it’s unclear what the well-posedness argument actually buys. This is consequential because it undermines the paper’s central motivation. The theoretical framing promises more than the method delivers.

Weakness 2: The experiments don’t adequately support the boundary condition claims. The experiments don’t clearly test the setting that motivates the paper. The strongest argument for giving boundary conditions as input is the time-dependent case. However, Appendix C only describes boundary values and fluxes sampled uniformly from [-10, 10], with no mention of time variation. The boundary experiments are also limited to the heat and advection equations with Dirichlet, Neumann and periodic conditions, even though Figure 1 also shows nonlocal conditions. In addition, the paper doesn’t compare against BOON (Saad et al., 2023), the most directly relevant prior work on enforcing boundary conditions in neural operators. It also describes BOON incorrectly. BOON enforces Dirichlet and periodic boundary conditions exactly, and Neumann conditions are enforced using a finite difference scheme. Appendix O instead claims that only Dirichlet is enforced exactly and is inconsistent about which boundary conditions BOON covers. This is consequential because the boundary transfer operator is one of the main contributions. Without a clear test on time-dependent boundary conditions or a comparison to BOON, it’s not possible to judge whether the proposed approach actually improves on existing methods in the setting the paper is designed for.

Questions for authors
  1. Problem statement and higher-order PDEs. The problem statement in Section 3 is restricted to first-order in time PDEs, but Section 4.1 claims the approach generalizes to higher-order temporal PDEs by adding initial snapshots. Could you generalize the problem statement accordingly? Have you tested this on any higher-order PDE?
  2. Admissible boundary operators. What exactly does “admissible boundary operators” mean? Please specify the function spaces for the boundary data, any compatibility conditions with the initial condition, and whether admissibility means the resulting problem satisfies Assumptions Q.1 and Q.2.
  3. Boundary conditions as input vs. satisfaction. The paper argues that a well-posed operator must take the boundary conditions as input, but the model doesn’t satisfy them exactly (Appendix O). If the boundary conditions can still be violated, what does the well-posedness argument actually provide? Why should they be inputs if they’re not enforced?
  4. PDE parameters as input. The motivation for including the boundary conditions is clear, especially for time-dependent ones. What is the motivation for including the PDE parameters? In many practical settings, the parameters may not be known exactly.
  5. Time-dependent boundary conditions. Were time-dependent boundary conditions tested? Appendix C only describes boundary values and fluxes sampled from [-10, 10], but Section 5.2 mentions the “temporal smoothness of the boundary data”. Since this is the main motivating case, please clarify and, if possible, add experiments on it. Similarly, Figure 1 shows nonlocal boundary conditions, but none appear in the experiments.
  6. Proposition 4.1. The proposition essentially states that the same input can’t map to two different outputs. This seems too trivial to be presented as a theoretical contribution. Can you clarify what it adds beyond this observation?
  7. Burgers split. For Burgers, training uses ν∈[5×10−4,5×10−2]\nu \in [5 \times 10^{-4}, 5 \times 10^{-2}] while testing uses ν∈[0.01,1.0]\nu \in [0.01, 1.0]. Most of the test range is outside the training range, so this is extrapolation rather than interpolation. Is this intended? If so, how does the model perform on test values within the training range?
  8. Navier-Stokes setup. Appendix C gives the kinematic viscosity as ν=ρuLRe\nu = \frac{\rho u L}{Re}, but it should be ν=ULRe\nu = \frac{UL}{Re}. The text also refers to “kinematic viscosity ranging from 100 to 5000”, which seems to be the Reynolds number range. Please clarify which quantity is varied.
  9. Data generation details. Which numerical solvers were used to generate the datasets? What time steps were used, and how were the Dirichlet and Neumann boundary conditions discretized? This is important for reproducibility.
  10. Initial conditions for Burgers. The Burgers datasets use initial conditions sampled from a Gaussian random field. Burgers equation models the transport of a quantity, and physically meaningful initial conditions are typically step functions (Riemann problems) or other structured profiles that lead to shocks and rarefaction waves. Gaussian random fields are commonly used in ML benchmarks, but they aren’t representative of the physical settings Burgers equation describes. How does the model perform on more physical initial conditions such as Riemann problems? Please also specify the covariance of the Gaussian random field used for all datasets.
3 comments

We thank you for your insightful comments on improving our paper. We hope the following could address your concerns.

W1, Q3 Boundary Conditions as Input vs. Satisfaction

Our well-posedness argument is not about the exact constraint satisfaction of the output. Rather, we contend that explicit input of PDE parameters and boundary conditions is necessary to completely specify the problem and uniquely determine the target solution. Without this information, the model must make averaged predictions or infer these conditions from initial states, which introduces irreducible error. The proposition makes no claim that predictions satisfy boundary conditions exactly; instead, what explicit conditioning provides is accuracy. As demonstrated in Tables 3, 14, and 15, explicit conditioning achieves higher accuracy both across the entire field and at boundary points.


W2 Comparison with BOON

BOON is not directly comparable because it addresses a different task. We study generalization of a single model across boundary conditions. BOON enforces a prescribed boundary condition through modifications to the operator kernel that are specific to each boundary type, which prevents a single model from generalizing across boundary types. Appendix O does not describe BOON incorrectly; it distinguishes two metrics. Boundary residual measures how well a prediction satisfies the boundary condition, while boundary nMSE measures the error of the predicted boundary values against the ground truth. Lines 1327–1329 state that BOON can achieve zero boundary residual. Lines 1304–1311 state that it achieves zero boundary nMSE only for Dirichlet conditions. BOON can have zero boundary residual but nonzero boundary nMSE because only Dirichlet conditions prescribe the boundary value itself.

Let uu be the true solution and u^i\hat{u}_i be the predicted value at grid node xix_i, with x0x_0 on the boundary. For a Dirichlet condition u(x0,t)=gD(x0,t)u(x_0, t) = g_D(x_0, t), BOON sets u^0=gD(x0,t)\hat{u}_0 = g_D(x_0, t), which is the exact boundary value, so the boundary error is zero. For a Neumann condition ∂u∂n(x0,t)=gN(x0,t)\frac{\partial u}{\partial n}(x_0, t) = g_N(x_0, t), BOON enforces a one-sided finite-difference stencil with coefficients aia_i by setting

u^0=1a0(gN(x0,t)−∑i≥1ai,u^i)\hat{u}_0 = \frac{1}{a_0} ( g_N(x_0, t) - \sum_{i \geq 1} a_i , \hat{u}_i )

This correction depends on the interior values u^i\hat{u}_i, which are the model's own predictions and carry irreducible error, so the corrected boundary value inherits that error. For periodic conditions, BOON enforces u^(x0,t)=u^(xN−1,t)\hat{u}(x_0, t) = \hat{u}(x_{N-1}, t), which still holds when both values share an error ee, i.e., u^(x0,t)=u^(xN−1,t)=u(x0,t)+e\hat{u}(x_0, t) = \hat{u}(x_{N-1}, t) = u(x_0, t) + e. A zero boundary residual therefore does not imply zero boundary nMSE.


Q1 Higher Order PDEs

For higher-order temporal PDEs, the initial condition simply contains more frames. For a PDE of order mm in time, the initial condition becomes the tuple u0=(u0,0,…,u0,m−1)\mathbf{u}_0 = (u_{0,0}, \dots, u_{0,m-1}), which in practice is the first mm frames of the trajectory. The solution operator and Proposition 4.1 hold unchanged, with XinitX_{init} replaced by the space of these tuples. We focus on first-order PDEs because this is the most common setting tested in existing benchmarks. Further, no existing benchmark dataset [2,3,4,5] contains higher-order PDEs with diverse boundary condition variations. Extensive benchmarking on higher-order PDEs therefore requires a community effort to first create standardized benchmark datasets, which is beyond the scope of this work and left for future work.


Q2 Admissible Boundary Operators

An admissible boundary operator is one for which the numerical solver used to generate the data has a solution and is unique. The boundary data lie in the space of bounded functions on the boundary over time, L∞(∂Ω×(0,T))L^\infty(\partial\Omega \times (0, T)). The neural operator is trained on and evaluated against trajectories produced by the numerical solver, so the operator it approximates is the solution map of the solver. Assumptions Q.1 and Q.2 therefore apply to this map.


Q4 PDE Parameters as Input

The motivation for including PDE parameters follows the well-posedness argument as for boundary conditions. By Proposition 4.1, an identical initial condition leads to different trajectories under different parameters. Without PDE parameters, the target solution is therefore not uniquely determined, and the model must either predict an average over parameters or infer it from initial frames. Further, like the majority of neural operators [5,6], we frame our model as a faster alternative to traditional numerical solvers. Since numerical solvers require the PDE parameters and boundary conditions as explicit inputs, it is fair to assume these inputs are known in this setting.

W2, Q5 Time Dependent Boundary Conditions

Our datasets do contain time-dependent boundary conditions, and it is sampled at every frame from the uniform distribution. By temporal smoothness in Section 5.2, we referred to the bounded variation of these values, which change at every frame but remain within a fixed range rather than varying by extremely large amounts between frames. The experiments evaluate the setting that motivates the paper. In each of them, explicit conditioning achieves higher accuracy both across the entire field and at boundary points.


W2, Q5 Limiting Boundary Conditions to Dirichlet, Neumann, and Periodic

We focus on Dirichlet, Neumann, and periodic conditions because they are the three most common boundary conditions studied in existing work [1,7], and they describe many real-world phenomena. Examples include fixed-temperature and no-slip walls (Dirichlet), insulated surfaces and prescribed heat flux (Neumann), and homogeneous turbulence and the longitudinal direction of global atmospheric models (periodic).

Additionally, generating data for a single PDE under Dirichlet, Neumann, and periodic boundary conditions cannot be done by switching a setting in one solver, because each boundary type calls for a different numerical discretization. The 2D viscous Burgers equation is a clear example. For periodic boundaries, a Fourier pseudo-spectral method is the natural choice, since Fourier modes satisfy periodicity automatically and give spectrally accurate derivatives. Fourier modes cannot represent walls, however, and using them on a domain with walls produces spurious oscillations at the boundary. Walls are instead handled naturally by the finite element method, where a Dirichlet condition is imposed by fixing the solution at the wall nodes and a Neumann condition enters the weak form as a boundary integral. One PDE with three boundary types therefore requires building, stabilizing, and verifying separate solvers. They must also be designed to produce data on the same grid and at comparable accuracy, so that the boundary condition is the only thing that changes across datasets. (We added experiments on Burgers’ equation with both PDE parameter and boundary condition variations to answer Reviewer 281A’s question.)

Existing benchmarks do not provide this kind of controlled variation. In widely used neural PDE solver benchmarks [2,3,4,5], each PDE comes with one boundary setting. None of them offers a single PDE under all three of Dirichlet, Neumann, and periodic boundary conditions. Building a boundary condition study from these benchmarks would mean taking each boundary type from a different PDE. The governing equation, the parameter range, and the resolution would then all change together with the boundary condition, and any difference in accuracy would be confounded rather than attributable to the boundary condition. For this reason we generated all datasets with boundary condition variation ourselves. Each of them required careful solver analysis and verification to guarantee high-fidelity data.

Because of the tremendous effort in writing numerical solvers, large-scale datasets with many different boundary condition variations for the same PDE call for a sustained community effort.


Q6 Proposition 4.1

Proposition 4.1 is rooted in classical PDE theory from Hadamard well-posedness. Our contribution is to formulate this requirement in the neural operator setting, which yields implications beyond the statement itself. First, the initial condition alone is insufficient to predict the future trajectory. Yet for time-dependent PDEs, neural operators are commonly formulated as a mapping from initial conditions to solution trajectories. Proposition 4.1 implies that this domain must be augmented to include the PDE parameter and boundary condition space. Second, any model without these explicit inputs must infer them from the initial states, which introduces irreducible error. These implications directly motivate our explicit conditioning on PDE parameters and boundary conditions.

Q7 Burgers Split

The larger testing range is intended. Following standard practice, the test range covers both in-distribution and out-of-distribution values. Below, we report the results on test values within the training range.

Method
ViT-27.29×10−27.29 \times 10^{-2}
ViT-58.44×10−38.44 \times 10^{-3}
ViT-106.57×10−26.57 \times 10^{-2}
Concat1.59×10−21.59 \times 10^{-2}
CAPE+Unet7.24×10−37.24 \times 10^{-3}
CAPE+FNO4.22×10−34.22 \times 10^{-3}
CAPE+Attn8.35×10−38.35 \times 10^{-3}
Unisolver6.31×10−36.31 \times 10^{-3}
MoE-POT3.70×10−23.70 \times 10^{-2}
Ours8.93×10−4\mathbf{8.93 \times 10^{-4}}

Q8 Navier-Stokes Setup

We thank the reviewer for pointing out this typo. The kinematic viscosity is ν=UL/Re\nu = UL/Re, where UU is the characteristic velocity and LL is the characteristic length. The varied quantity is the Reynolds number Re∈[100,5000]Re \in [100, 5000].


Q9 Data generation Details

We use different numerical solvers depending on the boundary condition type. For periodic boundary conditions, we use Fourier pseudo-spectral methods, which handle periodicity naturally. We use the 2/3 dealiasing rule and Crank–Nicolson time stepping for the incompressible Navier–Stokes equations, and with fourth-order Runge–Kutta (RK4) time stepping for the Burgers, heat, and advection equations. For Dirichlet boundary conditions, we use second-order central differences with Crank–Nicolson time stepping for the heat equation, and a second-order upwind scheme with explicit Runge–Kutta time stepping for the advection equation. Neumann conditions use the same schemes, with the prescribed fluxes imposed through ghost points. All solvers use adaptive time stepping governed by the CFL condition. This dependence of the solver on the boundary condition type is precisely why existing benchmarks typically fix a single boundary condition type for each PDE. Switching from periodic to Dirichlet or Neumann conditions requires replacing the solver entirely, not merely changing its inputs. It is also why building standardized benchmarks with diverse boundary conditions for the same PDE is challenging and will require long-term community effort.


Q10 Initial conditions for Burgers

Gaussian random fields have covariance operator C=σ2(−Δ+τ2I)−αC = \sigma^2 \left(-\Delta + \tau^2 I\right)^{-\alpha}, where Δ\Delta is the Laplacian, II is the identity, α\alpha controls the smoothness, τ\tau is the inverse length scale, and σ\sigma sets the amplitude. We use α=2.0\alpha=2.0, τ=1\tau=1, and σ=2\sigma=2.

Below, we report results on Riemann problems for the Burgers equation.

Method
ViT-21.13×10−11.13 \times 10^{-1}
ViT-51.79×10−21.79 \times 10^{-2}
ViT-105.61×10−25.61 \times 10^{-2}
Concat7.53×10−27.53 \times 10^{-2}
CAPE+Unet7.32×10−37.32 \times 10^{-3}
CAPE+FNO8.17×10−38.17 \times 10^{-3}
CAPE+Attn1.43×10−21.43 \times 10^{-2}
Unisolver2.81×10−22.81 \times 10^{-2}
MoE-POT3.15×10−23.15 \times 10^{-2}
Ours5.65×10−3\mathbf{5.65 \times 10^{-3}}

Thank you for your constructive feedback. We have updated the manuscript to incorporate your suggestions. We hope this clears up any confusion and concerns, and if so, we would greatly appreciate your consideration in raising the score.

[1] Guiding continuous operator learning through Physics-based boundary constraints, ICLR 2023

[2] PDEBENCH: An Extensive Benchmark for Scientific Machine Learning, NeurIPS 2022

[3] The Well: a Large-Scale Collection of Diverse Physics Simulations for Machine Learning, NeurIPS 2024

[4] FD-Bench: A Modular and Fair Benchmark for Data-driven Fluid Simulation, KDD 2026

[5] APEBench: A Benchmark for Autoregressive Neural Emulators of PDEs, NeurIPS 2024

[6] Weak baselines and reporting biases lead to overoptimism in machine learning for fluid-related partial differential equations, Nature Machine Intelligence

[7] Guiding continuous operator learning through Physics-based boundary constraints, ICLR 2023

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.