acceptodds
Under review as a conference paper at ICLR 2027

Preconditioned Physics-Informed Neural Operator Training

Abstract

Neural operators are typically trained in a supervised fashion, which requires a dataset to be generated with a classical solver. Training them physics-informed, i.e., purely from the governing equations, removes this large offline cost and allows fresh samples to be drawn at every optimization step, but has so far been limited to simplified problems and trails supervised training in accuracy. The obstacle is the ill-conditioning of physics-informed losses, which differential operators induce and which worsens as the discretization is refined. We therefore propose a preconditioned residual loss function and show mesh-independent conditioning for elliptic problems and greatly improved conditioning for saddle point problems. Realized through geometric and algebraic multigrid, the construction applies to linear and nonlinear equations, steady or time-dependent, on structured and unstructured meshes, is agnostic to the neural operator architecture, and adds no cost at inference. On the Poisson, Allen-Cahn and stationary Stokes equations, the resulting label-free training matches supervised training and is four to twenty-five times more accurate than previous physics-informed operator learning methods.

Primary area:
optimization
Keywords:
Neural Operators, Physics-Informed Learning, Preconditioning

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.25 over 4 reviews, above 96% of the others.
Reviewer_G7ToOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This manuscript proposes a preconditioned residual loss function and preconditioned PDE solver training routine, realized through geometric and algebraic multigrid. The proposed routine applies to linear and nonlinear equations, steady or time-dependent, on structured and unstructured meshes, is agnostic to the neural operator architecture, and adds no cost at inference.

Critical strengths
  • Strong motivation and a clear mathematical interpretation of the proposed components.
  • Seamless integration with classical and neural PDE-solving methods, with many favorable properties (such as not requiring changes to the NO architecture, applying to multiple PDE formulations across multiple discretization routines, etc.)
  • Theoretical analysis of how preconditioning changes the conditioning of the residual loss.
  • Performance gain across multiple baselines and PDEs.
Critical weaknesses
  • Improving the conditioning of the middle PDE-dependent factor does not necessarily guarantee good conditioning of the full parameter-space matrix. This is also an author-recognized theoretical limitation.
  • Performance gains over "supervised learning" routines on Poisson and Allen-Cahn are less significant than on Stokes, which limits the practical scope of this solver, especially in cases that do not need to be physics-informed.
  • Baselines only cover PINO and physics-informed DeepONet, which are old baselines. No newer baseline is discussed or compared with.
  • The proposed nonlinear preconditioning strategy is problem-specific and still needs manual design/tuning. The overall process is not automated, and adapting the solver to new problems requires manual effort and a solid mathematical background. The actual generalization cost is higher than claimed in the paper.
Questions for authors
  • Please see "weaknesses" above.
2 comments

We thank you for taking the time to review our manuscript. We comment on your mentioned weaknesses:

  1. The hypothesis of the manuscript is that supervised neural operator learning works ``reasonably well'', especially in comparison to the physics-informed alternatives before this work. In that sense our goal was to mitigate the spectral ill-conditioning introduced through physics-informed loss functions. To which extend ill-conditioning of neural networks in general is a problem is a broader question that we do not attempt to answer. Historically, the introduction of initialization schemes and momentum-based first order methods were landmark results in this direction. We see our contribution as orthogonal to that development.

  2. We don't expect to improve upon data-driven learning, even though physics-informed learning improved upon Stokes in our specific test example. Rather, the purpose of the method is to eliminate the data-generation process and allow training at the infinite-data limit where no sample is revisited.

  3. We now consider more baselines, including a preconditioned variant of DeepONet and Transolver, please compare to the global answer for details.

  4. We agree with this statement. Preconditioning as a mathematical discipline however is mature, so for many equations a suitable preconditioner is textbook material. We therefore see the reviewer's observation as a strength as well: physics-informed operator learning can draw directly on decades of preconditioning research for classical PDE solvers, and every advance there carries over without changes to the network architecture.

Thank you for your detailed explanation. I am keeping my original score of weak accept.

Reviewer_k3x7OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper addresses ill-conditioning induced by discretized PDE operators in physics-informed neural operator training. It proposes a loss that preconditions finite element residuals: geometric multigrid approximate inverses are used for Poisson and Allen–Cahn, while block-diagonal residual weighting incorporating algebraic multigrid is used for Stokes. Preconditioning is confined to the training loss and adds no preconditioning computation at inference. The authors analyze conditioning in the discrete output space and evaluate the approach using FNO and GAOT. The reported results approach supervised accuracy on Poisson, retain an accuracy gap on Allen–Cahn, and outperform supervised training on Stokes. Comparisons with unpreconditioned residual losses and ablations of preconditioning strength provide further support for the approach.

Critical strengths
  1. The method targets a clear numerical bottleneck and is practically applicable. The authors apply classical preconditioning to discrete residual losses for neural operators, mitigating ill-conditioning introduced by the PDE operator. The construction is compatible with different network backbones and adds no preconditioning computation at inference. The linear-model experiment in Figure 1 and the preconditioning-strength sweep in Figure 2 provide direct support for the proposed mechanism.

  2. The empirical gains under the reported fixed training budgets are persuasive. In Table 1, the Poisson error is 5.1±0.8%5.1\pm0.8\%, close to the supervised result of 4.7±0.8%4.7\pm0.8\%. Stokes velocity and pressure errors are 1.6±0.2%1.6\pm0.2\% and 0.85±0.04%0.85\pm0.04\%, compared with supervised errors of 3.3±0.2%3.3\pm0.2\% and 1.79±0.16%1.79\pm0.16\%. Comparisons with unpreconditioned residual training using the same backbone in Figure 5 and Tables 4 and 6 further support attributing the gains to preconditioning rather than solely to architectural differences. Here, the reported uncertainties are half-ranges across three seeds.

Critical weaknesses
  1. The experimental pipeline’s dependence on reference solutions needs clarification. Appendix E, page 21, states that each Stokes sample is normalized using the finite element L2L^2 norm of its reference velocity, so constructing the experimental input distribution uses reference solutions. Hyperparameter search and checkpoint selection also use labeled validation errors. Labeled validation is a reasonable experimental protocol and does not contradict the absence of training-solution targets in the residual loss. However, these dependencies should be distinguished from eliminating reference-solution generation throughout the pipeline. This distinction matters for the claimed data-generation cost advantage. A normalization ablation that does not require reference solutions, together with explicit cost accounting, would strengthen this claim.

  2. The theoretical claims require more precise assumptions. On page 4, preservation of residual zeros is used to claim that preconditioning does not change the minimizers. For a restricted neural-network class that cannot achieve zero residual, changing the residual norm can change the optimal parameters. Moreover, although the condition-number inequalities in Lemma 1 are valid, the uniform Euclidean condition-number bounds needed for the actual multigrid construction require specific justification; eigenvalue clustering alone does not provide singular-value bounds. Figure 6 evaluates resolution transfer rather than conditioning and complexity as the training mesh is refined. These gaps limit the strength of the theoretical claims without directly invalidating the reported empirical gains.

Questions for authors
  1. For Stokes, does the relative performance of preconditioned and supervised training persist when normalization depends only on the input or uses a fixed global scale? Please clarify how reference solutions enter preprocessing, hyperparameter search, and model selection, and specify the intended scope of “label-free.”

  2. Does the statement on page 4 that preconditioning preserves minimizers assume that the network can achieve zero residual? In the non-realizable case, should this statement be restricted to preservation of the residual-zero set, acknowledging that different norms may yield different best approximations?

  3. For the implemented multigrid V-cycle and approximate Stokes blocks, can you provide uniform bounds in the relevant norms, or precise theoretical references, supporting the stated condition-number scaling? If the theoretical assumptions do not cover the implementation, could you report conditioning and time to reach a common error target across multiple training meshes to substantiate the existing claims?

2 comments

Thank you for the review. We answer your questions below.

  1. This point has been raised by several reviewers and we address it in the global answer.

  2. We thank you for catching this imprecise statement. Yes, as stated it needs the network to reach zero residual. We will state this more precisely. In the non-realizable case minimizers will differ. Additionally, optimization dynamics differ, too.

  3. For Poisson: We thank you for catching this imprecise statement. Standard GMG theory gives a mesh-independent contraction ∣I−PA∣A≤δ<1\\|I-PA\\|_A \le \delta < 1 in the energy norm, i.e.uniform bounds on the eigenvalues of PAPA, not on the Euclidean condition number κ2(PA)\kappa_2(PA). Therefore, we measured the Euclidean condition number κ(PA)\kappa(PA) empirically for the considered examples, also for a refining sequence of meshes:

Grid17×1717\times1733×3333\times3365×6565\times65129×129129\times129257×257257\times257
κ2(PA)\kappa_2(PA)1.081.081.121.121.141.141.151.151.161.16

We will clarify this distinction in the revised manuscript.

For Stokes: Here Lemma~1(2) with B=PSB = P_S suffices, so eigenvalue bounds are enough at the price of κ(PS)=O(h−2)\kappa(P_S) = \mathcal O(h^{-2}), which we state in the manuscript. The textbook result for this eigenvalue clustering is Murphy et al [A note on preconditioning for indefinite linear systems] and is cited in the manuscript. For the approximate version that we use (one AMG V-cycle for KK, lumped pressure mass for SS, ωS=16\omega_S = 16) we computed the condition numbers under mesh refinement, and compare to the exact blocks:

Mean edge hhFree DOFsNo precond.: κ(A2)\kappa(A^2)AMG block: κ(APSA)\kappa(AP_SA)Exact block: κ(APSA)\kappa(AP_SA)
0.0570.0572 8772\,8771.1⋅10101.1\cdot10^{10}4.6⋅1054.6\cdot10^{5}4.6⋅1054.6\cdot10^{5}
0.0480.0484 0744\,0742.2⋅10102.2\cdot10^{10}6.0⋅1056.0\cdot10^{5}6.0⋅1056.0\cdot10^{5}
0.0390.0396 1896\,1895.1⋅10105.1\cdot10^{10}9.3⋅1059.3\cdot10^{5}9.0⋅1059.0\cdot10^{5}
0.0340.034 (paper)8 4638\,4631.2⋅10111.2\cdot10^{11}1.3⋅1061.3\cdot10^{6}1.3⋅1061.3\cdot10^{6}
0.0290.02911 43311\,4331.9⋅10111.9\cdot10^{11}1.7⋅1061.7\cdot10^{6}1.7⋅1061.7\cdot10^{6}
0.0250.02515 96915\,9693.5⋅10113.5\cdot10^{11}2.5⋅1062.5\cdot10^{6}2.4⋅1062.4\cdot10^{6}
Fitted rateh−4.13h^{-4.13}h−2.01h^{-2.01}h−1.96h^{-1.96}

The block preconditioner reduces the growth from h−4h^{-4} to h−2h^{-2}, as predicted, and the AMG block stays within 6%6\% of the exact block at every mesh.

Thank you for the detailed response and the additional experiments in the Global Answer. The residual-based model selection, mesh-refinement study, and conditioning measurements substantially address my main concerns. I also appreciate the clarification regarding minimizer preservation. Please incorporate the realizability assumption, clarify the theoretical and empirical support for the conditioning claims, and explicitly document the 16 calibration solves used for Stokes normalization and their cost. The training-overhead claims should also reflect the updated timing results. Overall, I maintain my original rating.

Reviewer_8yEeOpenReview ↗
Rating
4(+2)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposed a preconditioned physics informed Loss for neural operator training, with GMG or AMG as preconditioner. Experiments show that the proposed Loss is easier to optimize compared to Least Square Loss. The manuscript is overall clearly written.

Critical strengths
  1. The main idea is easy to follow.
  2. Positive empirical performance on the controlled benchmarks.
Critical weaknesses
  1. The experiments are conducted on small-scale test instances
  2. Too few baseline networks are included in the comparison.
  3. Some existing works have already explored similar ideas.
Questions for authors

Q1 Scalability of training resolutions.

the reported data resolutions is too small (65×6565\times 65 or 129×129129\times 129), please provide results on larger training grids to assess whether the accuracy benefits persist as the resolution increases. (for example 256×256256 \times 256 and 512×512512 \times 512)

Q2 why the additional training cost is negligible?

line 115-116, the author said "...adds negligible cost during training...", but in fact the preconditioner used here is GMG or AMG, which need to take a forward and backward pass during each training step, I think it's should be expensive...especially with the larger grid size.

Q3 limited backbone evaluation.

Since the paper's main contribution is to propose a new type of training loss, it is important to demonstrate the generality of this loss. However, the authors only evaluate its effectiveness with FNO and GAOT. Please test additional architectures, such as DeepONet, U-Net, MgNO, Transolver, and GNO.

Q4 Some existing works need to check.

These works share similar ideas with this paper, and the authors need to distinguish their contribution from these works:

[1] Vescovini et al. Preconditioned Variational Physics-informed Neural Operator (P-VINO).

[2] Liu et al. Preconditioning for Physics-Informed Neural Networks.

[3] Wang et al. Born-Preconditioned Learning for Iterative Neural PDE Solvers.

5 comments

We thank the reviewer for the thorough assessment of our manuscript.

Q1 Scalability of training resolutions Please refer to the global answer.

Q2 Why is the additional training cost negligible? For one training step (forward, loss, backward; batch 32; RTX 4090) the overhead of LPLSL_\textup{PLS} over LLSL_\textup{LS} (hence the cost of GMG) is +172+172\\%, +67+67\\%, +18+18\\% and +6+6\\% at 33233^2, 65265^2, 1292129^2 and 2572257^2 and similarly measured against LdataL_\textup{data} (+179+179\\% down to +8.5+8.5\\%). GMG is cheap as a V-cycle performs a fixed number of smoothing sweeps and grid transfers per level, and each coarser level has a quarter of the nodes. Its complexity is O(n)\mathcal O(n) when nn is the number of degrees of freedom in the mesh, comparable to evaluating the residual once, and far below a forward and backward pass of the network. On a GPU a V-cycle is a short sequence of small kernels whose run time is dominated by launch latency. The relative overhead is therefore large on coarse grids and small on the fine grids. With algebraic multigrid a step is 55--99\\% more expensive than with GMG (24.624.6 vs. 23.323.3ms at 33233^2, 465465 vs. 456456ms at 2572257^2).

Q3 Limited backbone evaluation Please refer to the global answer.

Q4 Additional Literature Thank you for the pointer to the literature. We will discuss the references in the revised manuscript. We give a brief discussion already here in how these works differ from ours: [1] This paper considers energy formulations only and additionally makes the preconditioner a part of the neural operator (a last non-trainable layer) so it needs to be evaluated at inference time. Our work on the other hand is not restricted to energy formulations and does not require the preconditioner to be evaluated at inference time. Note that not all PDEs possess an energy formulation. [2] This work considers PINNs and uses ILU as a preconditioner, whereas we work with neural operators and multigrid. [3] In this work the preconditioner is a learned object, which is rather different from our approach

Thanks, I have decided to raise my score to Weak Accept.

I have one additional question regarding the potential extension of your framework.

Could the proposed preconditioned physics-informed loss also be applied to learning neural preconditioners, rather than solution operators?

For example, one could use a classical multigrid (MG) preconditioner as a baseline and train a neural operator to learn an additional correction, similar in spirit to Born-Preconditioned Learning (~[3] Wang et al). In this setting, MG could potentially serve both as a baseline preconditioner and as part of the training loss.

Do you think your framework could naturally extend to this setting? In particular, would the preconditioned loss provide advantages when the learned operator is subsequently used as a preconditioner within an iterative solver?

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

The paper proposes a way to train neural operators to solve PDEs without requiring ground-truth PDE solutions as training labels, while achieving accuracy comparable to supervised learning. The method is evaluated on three PDE problems: Poisson, Allen–Cahn, and stationary Stokes equations. The experiments include both structured and unstructured meshes, using FNO and GAOT as neural operator backbones. The authors compare their approach with supervised training, standard residual-based training, PINO, and physics-informed DeepONet. The proposed loss substantially improves convergence and prediction accuracy over existing physics-informed methods.

Critical strengths

1.The paper identifies poor conditioning as a key reason why physics-informed training struggles. I found the analysis in Section 2 and the controlled experiment in Figure 2 convincing.

  1. The improvements over unpreconditioned training are substantial, especially for Poisson (Figure 2) and Stokes (Figure 5 and Table 6). I also appreciate that the experiments cover a nonlinear time-dependent problem (Figure 4) and an unstructured-mesh saddle-point problem (Figure 5), rather than only standard Poisson equations. The Stokes results are particularly encouraging, as the proposed method outperforms supervised training on both velocity (1.6% vs. 3.3%) and pressure (0.85% vs. 1.79%), as reported in Table 1.
Critical weaknesses
  1. Although the paper provides theoretical results on mesh-independent conditioning for elliptic problems, the experiments do not directly evaluate how training performance scales with mesh resolution. Figure 6 in Appendix A studies resolution transfer, but the model is trained at a fixed resolution of (65\times65). I would like to see a mesh refinement experiment where separate models are trained at different resolutions, comparing convergence, prediction accuracy, and computational cost. This would provide more direct evidence for the practical benefit of the proposed conditioning analysis.
  2. One of the main motivations of this work is to avoid generating ground-truth PDE solutions. However, the experimental protocol still relies on labeled validation data for hyperparameter tuning and checkpoint selection (Appendix C, Table 3; Appendix E). Although the training loss itself is label-free, it is unclear whether similar performance can be achieved without access to reference solutions during model selection. I would like to see an experiment using only physics-based validation criteria, with a comparison to the current results.
Questions for authors

Could the authors provide a mesh refinement experiment where models are independently trained at different resolutions (e.g., (33\times33), (65\times65), (129\times129), and (257\times257))? How do convergence, prediction accuracy, and computational cost change as the resolution increases? This would help verify whether the theoretical improvement in conditioning translates into more efficient neural operator training at finer resolutions.

Is it possible to perform hyperparameter tuning and checkpoint selection using only physics-based validation criteria, without access to ground-truth PDE solutions? If so, how does the resulting test accuracy compare with the current results obtained using labeled validation data? There are a few minor typos in the paper: “in a unsupervised fashion” should be “in an unsupervised fashion” (p. 1, Introduction); “an coordinate-based formulation” should be “a coordinate-based formulation” (p. 4, Section 2.1); “Appendix 23” appears to refer to Appendix D (p. 7, Section 3.2); and “This is happens also” should be “This also happens” (p. 13, Appendix A).

1 comment

We thank the reviewer for raising valid points and we have performed two additional experiments as requested: (i) a mesh refinement study for the Poisson equation and (ii) a training pipeline that is fully physics-informed and does not rely on labeled data for hyperparameter selection. Both can be found in the global answer, as these were points raised by several reviewers.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Global Answer Part I

We answer here questions that have been raised by more than one reviewer: a mesh refinement study, a fully simulation free pipeline, and additional baseline evaluations.

Mesh Refinement

We investigate scalability with respect to the mesh size for the Poisson equation on the dataset employed in the paper in Table 1 (the benchmark table). We sweep the grid from 33×3333\times 33 to 257×257257\times 257 and recorded the results below:

GridRel. test L2L^2 error [%], LdataL_\text{data}Rel. test L2L^2 error [%], LPLSL_\text{PLS}Rel. test L2L^2 error [%], LLSL_\text{LS}Epochs to 10 % val. error (LdataL_\text{data} / LPLSL_\text{PLS})Time per epoch [s] (LdataL_\text{data} / LPLSL_\text{PLS})
33×3333\times334.50±0.634.50 \pm 0.634.92±0.544.92 \pm 0.5434.8±1.034.8 \pm 1.0145 / 1580.24 / 0.67
65×6565\times654.74±0.824.74 \pm 0.825.11±0.855.11 \pm 0.8546.6±1.046.6 \pm 1.0151 / 1710.37 / 0.83
129×129129\times1295.01±0.985.01 \pm 0.985.29±0.955.29 \pm 0.9565.2±0.565.2 \pm 0.5165 / 1792.40 / 2.92
257×257257\times2575.22±1.075.22 \pm 1.075.52±1.145.52 \pm 1.1475.8±0.775.8 \pm 0.7172 / 1878.44 / 9.16

Errors are mean ±\pm half-range over three seeds; epochs to 10 %10\,\% validation error are averaged over the seeds (LLSL_\textup{LS} does not reach 10 %10\,\% at any grid). The results show that

  1. LPLSL_{\text{PLS}} achieves a similar accuracy under mesh refinement compared to LdataL_{\text{data}} with at most 9% higher error (for the 33×3333\times 33 grid) and less for finer grids.
  2. The number of epochs to reach 10% validation error with LPLSL_{\textup{PLS}} does not deteriorate (measured relative to reaching it with LdataL_{\textup{data}}) as the mesh is refined.
  3. The overhead of LPLSL_{\textup{PLS}} shrinks from 179% at 33×3333\times 33 to 8.5% at 257×257257\times 257 as the mesh is refined. This is due to the dominating cost of launching GPU kernels at the small resolution, which is proportionally smaller for finer meshes, the details are in the table below.
Grid33×3333\times3365×6565\times65129×129129\times129257×257257\times257
Step with LLSL_\text{LS} [ms]8.68.624.124.1121.6121.6428.8428.8
Step with LPLSL_\text{PLS} (GMG) [ms]23.323.340.340.3143.1143.1456.3456.3
Overhead+172+172\\%+67+67\\%+18+18\\%+6+6\\%

A V-cycle costs 33--1010,ms at every grid, as it is bound by kernel-launch latency, so the overhead is large on coarse grids and small on fine ones. The revision uses these new, carefully timed numbers. Additionally, LLSL_{\textup{LS}} deteriorates in accuracy along the refinement which illustrates the growing ill-conditioning as the mesh is refined.

Fully Physics-Informed Pipeline

In the submission, labels entered (a) through learning rate tuning (and ωS\omega_S for Stokes) and checkpoint selection (App.~C, E); and (b) for Stokes, every sample, including the training samples, was rescaled so that its reference velocity has unit norm, which requires one solve per sample.

(a) Selection. Every run now also logs the preconditioned residual of the validation set, 12∣P(Auθ−b)∣2\tfrac12\\|P(A u_\theta - b)\\|^2, and keeps a second checkpoint selected by it alone. Test errors [%], three seeds:

Label-selectedResidual-selected
Poisson5.20±0.825.20 \pm 0.825.23±0.855.23 \pm 0.85
Allen–Cahn4.05±0.234.05 \pm 0.234.03±0.274.03 \pm 0.27
Stokes, uu / pp1.57±0.221.57 \pm 0.22 / 0.84±0.040.84 \pm 0.041.57±0.231.57 \pm 0.23 / 0.84±0.040.84 \pm 0.04

Across learning-rate sweeps the residual picks the same rate as the labels (Poisson 3⋅10−43\cdot10^{-4}, Allen--Cahn 10−310^{-3}). For ωS\omega_S, runs are compared in one fixed reference norm; with reference weight ≥4\geq 4 the choice lands on the same ωS∈[4,16]\omega_S \in [4,16] plateau as the labels.

(b) Stokes scaling. We replace the per-sample scaling by one computed from the input: ff is scaled by ∣f∣\\|f\\| times the mean ratio ∣u∣/∣f∣\\|u\\|/\\|f\\| of 16 calibration solves, done once for the data set. This improves both losses: LPLSL_\textup{PLS} 1.07±0.121.07 \pm 0.12 / 0.64±0.030.64 \pm 0.03 against LdataL_\textup{data} 2.30±0.102.30 \pm 0.10 / 1.27±0.061.27 \pm 0.06 (submission: 1.571.57 / 0.840.84 against 3.253.25 / 1.751.75).

Note: Allen--Cahn errors here are the trainer's space-time error over steps 1--10, not Table 1's metric (3.2 %).

AuthorsOpenReview ↗

Global Answer Part II

Limited Backbone Evaluation

Several reviewers required additional baselines. We added DeepONet and Transolver to the FNO (Poisson) and GAOT (Stokes) test cases of the paper. Relative test L2L^2 errors [%], mean ±\pm half-range over three seeds are below

BackboneLdataL_\text{data}LPLSL_\text{PLS}LLSL_\text{LS}
Poisson (65×6565\times65)
FNO (paper)4.7±0.84.7 \pm 0.85.1±0.85.1 \pm 0.846.6±1.046.6 \pm 1.0
DeepONet9.86±0.329.86 \pm 0.3210.07±0.2710.07 \pm 0.2746.1±2.546.1 \pm 2.5
Transolver2.88±0.352.88 \pm 0.352.76±0.302.76 \pm 0.3063.0±22.263.0 \pm 22.2
Stokes**, velocity / pressure**
GAOT (paper)3.3±0.23.3 \pm 0.2 / 1.79±0.161.79 \pm 0.161.6±0.21.6 \pm 0.2 / 0.85±0.040.85 \pm 0.04–
DeepONet27.97±0.7627.97 \pm 0.76 / 30.20±1.6130.20 \pm 1.6134.75±1.0434.75 \pm 1.04 / 30.90±0.8130.90 \pm 0.8172.0±2.872.0 \pm 2.8 / 42.6±1.342.6 \pm 1.3
Transolver6.33±0.536.33 \pm 0.53 / 5.89±1.265.89 \pm 1.268.66±1.398.66 \pm 1.39 / 4.48±0.834.48 \pm 0.8387.1±5.887.1 \pm 5.8 / 27.1±1.327.1 \pm 1.3

The additional backbones confirm that preconditioning is what makes physics-informed operator learning work, independently of the architecture: LLSL_\textup{LS} fails on every backbone, while LPLSL_\textup{PLS} closely tracks supervised training. On Poisson, LPLSL_\textup{PLS} is within 2% of LdataL_\textup{data} for DeepONet and even 4% more accurate for Transolver. On Stokes, it is within a factor of 1.41.4 of LdataL_\textup{data} in velocity and matches or improves on it in pressure. As for FNO and GAOT, the DeepONet and Transolver outputs are interpolated into the finite element space (Section~2), which is all that is needed to apply the preconditioner.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.