acceptodds
Under review as a conference paper at ICLR 2027

Amortized Physics-Informed Learning via Generative Initialization of Radial Basis Functions

Abstract

Physics-Informed Neural Networks (PINNs) commonly optimize a new solution model for each PDE instance, making repeated changes in parameters, boundary conditions, or forcing terms costly. GI-RBF addresses this limitation through a decoupled offline–online framework combining an explicit anisotropic Radial Basis Function (RBF) solver with a task-conditioned generative initializer. The initializer amortizes the search for the complete RBF state—including coefficients, centers, and kernel geometry—evaluated once per query; its output is then refined directly under the deployment physics objective. Analytic derivatives and the localized explicit representation enable inexpensive physics updates, while learned initialization reduces the refinement needed to reach useful accuracy. Experiments on four canonical PDE families and a transient -dimensional Flow Mixing problem show improved finite-budget adaptation over standard RBF initialization across interpolation and extrapolation regimes. The resulting inference-through-adaptation paradigm reaches an accurate solution through fast, physics-based refinement at deployment.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Physics-Informed Neural Networks (PINNs), Radial Basis Functions (RBFs), partial differential equations (PDEs), generative initialization, task-conditioned learning, offline–online framework, physics-based refinement, amortized inference, anisotropic kernels, analytic derivatives, finite-budget adaptation, interpolation, extrapolation, flow mixing, inference-through-adaptation

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_XCkEOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposed GI-RBF, a decoupled offline-online framework that uses Gaussian EBF to represent PDE solutions. The method includes a task conditioned generative initializer to predict the RBF parameter set which includes the coefficients, centers, scales, orientations and geometry. In the offline stage, the initializer is trained with solution supervision with short physics based optimization trajectories. In the online stage, its predictions are further refined through steps of physics-informed optimization. Experiments were conducted on four PDE families and a flow-mixing problem. With a matched online refinement budget, their method improves over PINN parameter methods and standard RBF initialization.

Critical strengths

1.The comparison with standard-RBF is well controlled. Both methods use the same setting so the effectiveness of learned initialization is established with controlled comparison.

2.The ablations are comprehensive and help to identify which components of the design play a major role in performance. The negative results on variational sampling and zero-shot self-bias are honest.

3.The learned initialization accelerates adaptation under limited online refinement budgets. This decoupling offline-online framework requires fewer optimization steps to achieve same accuracy.

4.The method uses a RBF solver that uses few trainable parameters and demonstrates a favorable accuracy-efficiency trade-off in comparison with MLP-PINN and SPINN in figure 2 and table 5.

Critical weaknesses

1.It is unclear whether all components in the framework contribute to the performance. For example in table 19 the result shows that in the single-start Allen-cahn experiment, the simplest deterministic initializer performs the best compared with stochastic and variational methods. Similarly in table 17, full-field zero-teacher control attains an effective initialization. While a 25-step teacher cannot consistently improve the endpoint medians across different confirmation panels.

2.The paper compared with baselines such as Hyper-LR-PINN, PINO, P2INN-SVD, standard RBF, however it is difficult to tell whether it is the gain brought by the RBF solver or the method’s generative initialization strategy. The authors might consider adding in learned initialization baselines as direct comparison such as Reptil/MAML style to showcase the advantage of generative initialization.

3.The end-to-end computational efficiency is not fully established. Table 16 shows that about 1000-1500 PDE queries are needed to offset the offline training cost relative to standard-RBF. The comparison with other learned PDE solvers such FNO and PINO are not included, and their offline training and deployment costs differ a lot. Given the paper claims efficient multi-query PDE solving, I recommend a more complete evaluation of total offline and online costs at fixed accuracy threshold to establish the claim.

Questions for authors

1.Can the authors evaluate the deterministic, no-KL initializer across all four PDE families with the same online refinement setting?

2.Can the author directly compare the time-to-target of refinement-informed teacher and direct conditional training to show the 25-step teacher component can accelerate adaptation?

3.How does the method compare with other shared meta-learned RBF initialization methods under the same online refinement? Or compared with other methods with a warm start from a neighboring training PDE condition.

4.The repeated query amortization in the main experiment and in table 16 have different configurations as one has 30, 000 outer updates one has 3,000 outer updates. Can the author report using the same fixed configurations and re-calculate the repeated-query crossover?

5.Could the author further clarify the methodological novelty of GI-RBF relative to current hypernetwork approaches such as HyperPINN and Gaussian-based solvers such as PIG. Given that conditional parameter generation and adaptive Gaussian representation have been explored in the literature, what is the distinct advantage of amortizing the initialization of the complete RBF parameter set?

6 comments

W1 and Q1 - Deterministic, variational, and teacher contributions

Thank you for asking us to clarify the roles of these two components. We agree that the main text did not state them clearly. The core of GI-RBF is the conditional initializer; both components are optional training choices on top of it:

  • Variational training. All reported results decode the conditional latent mean, i.e., one initialization per PDE. The same architecture can be trained deterministically, without the KL penalty and with the conditional mean instead of latent samples, and this deterministic initializer is sufficient for the single-start acceleration. Latent sampling adds the option of generating several candidate initializations for the same PDE; KL is not required for this and gives a bounded additional benefit (Appendix D.3).

  • Refinement-informed training. Direct conditional training already gives useful initializations (Table 17). The teacher additionally aligns the initializer with the deployment optimizer, which reduces the refinement steps needed to reach the target at a higher offline cost.

We clarified these roles in the main text (Section 4.3: page 5, lines 216–220; Section 4.4: pages 5–6, lines 267–272; Experiment II: page 7, lines 334–337). Below, we answer Q1 with evidence for both.

1. Deterministic initialization is sufficient across all four PDEs. Deterministic and variational initializers use the same refinement teacher, online refiner and training budget. The deterministic initializer has lower error after 50 refinement updates in all eight panels.

PDEDeterministic GI: error after 50 updates (%), I / EVariational GI: error after 50 updates (%), I / E
Burgers12.9 / 18.919.6 / 21.1
Convection2.1 / 4.124.1 / 21.2
Allen–Cahn0.9 / 2.31.5 / 4.2
Helmholtz60.8 / 55.1151.8 / 154.1

Rebuttal Table R3-T1. Mean relative-L2 error (%) after 50 refinement updates, Interpolation / Extrapolation; three training seeds. Both initializers use the same teacher, online refiner and training budget. Full results: Appendix D.3, Table 19a.

In the separate shared-initialization comparison, deterministic GI also reaches 5% error in fewer median steps than Standard-RBF in six of eight panels, and as often or more often in all eight (Appendix D.4, Table 19b).

2. Refinement-informed training reduces refinement steps, at higher offline cost. With the deterministic initializer and online refiner held fixed, adding the teacher reduces the median steps to 3% error in three of four interpolation/extrapolation panels across Burgers and Allen–Cahn, and matches the fourth, without reducing target reach. This requires 2.3–2.7× the offline training time.

Final error also improves on Burgers, but the effect is mixed on Allen–Cahn. Convection and Helmholtz were not tested in this ablation. The teacher is therefore a useful training option, rather than a requirement for effective initialization.

Full results, settings and per-seed detail: Appendix D.1, Table 17a.

W2 and Q2 - Direct initialization baselines

Thank you for suggesting this comparison. We added Reptile-RBF and Shared-GI to assess whether task conditioning improves adaptation beyond a shared learned initialization. Reptile-RBF learns one shared initialization through first-order updates toward task-adapted parameters, without differentiating through refinement. Shared-GI retains our deterministic generator and training objective but uses a fixed condition, providing a direct control for conditioning. These comparisons complement Standard-RBF (submitted Figure 3 and Table 8), which tests the benefit of learned versus standard initialization under identical refinement.

We have not evaluated nearby-condition warm starts. That comparison would require specifying how training-task solutions are stored and how the nearest condition is selected.

Comparison. We evaluate all three learned initializations across four PDE families and two seeds, using the same 1,536 online parameters, test cases, physics loss, AdamW settings and GPU. All three receive the same number of training updates (Appendix D.4, Table 19b). Conditional GI and Shared-GI also match per-seed training tasks, architecture and objective.

PDEMethodError after 25 updates (%), I / E ↓Reach ≤5% (%), I / E ↑
BurgersConditional GI14.1 / 19.985 / 100
BurgersShared-GI18.7 / 23.285 / 100
BurgersReptile18.8 / 19.985 / 100
ConvectionConditional GI2.6 / 6.5100 / 100
ConvectionShared-GI21.7 / 31.7100 / 100
ConvectionReptile10.6 / 14.4100 / 100
Allen–CahnConditional GI34.1 / 49.485 / 40
Allen–CahnShared-GI26.8 / 38.497.5 / 45
Allen–CahnReptile24.9 / 38.2100 / 45
HelmholtzConditional GI71.5 / 80.657.5 / 62.5
HelmholtzShared-GI126.2 / 136.830 / 25
HelmholtzReptile92.8 / 83.252.5 / 40

Rebuttal Table R3-T2. Error averages all cases after 25 refinement updates. Reach counts cases attaining ≤5% within 1,000 updates, including all 40 seed–case observations per panel.

Shared initialization is a strong baseline: all three methods reach the target equally often on Burgers, and the shared starts are better on Allen–Cahn. Conditioning, however, lowers early error relative to Shared-GI on Burgers, Convection and Helmholtz. The clearest case is Convection, where conditional GI reaches 5% error in a quarter to half of the updates the shared initializations need. On Helmholtz, it also reaches the target most often.

Overall, task conditioning improves on a shared learned initialization on three of the four equations.

Q3 - Consistent repeated-query accounting

The initial 30k recipe used a fixed training budget. This choice was rather conservative, which inflated the break-even point. The reduced 3k training was introduced to show that the break-even point is a direct function of the GI training, which can be further optimized and thus result in a more attractive break-even point. We therefore present transparent reporting of the break-even point and accuracy at the higher-end (30k) and lower-end (5k) budgets of the same recipe, where the only modification is the number of training updates. Each budget is one fixed configuration with its own measured training cost, and query time is measured directly, including generator inference (about 2.5 ms) and online refinement.

PDETraining updatesOffline cost (s)Saving/query vs Standard-RBF (ms)Error @1,000 (%)Reach ≤5% (%)Break-even queries
Burgers30k5,640188 / 1052.60 / 1.7685 / 10029,933 / 53,775
Burgers5k1,018181 / 932.82 / 1.8185 / 955,627 / 10,996
Convection30k3,863110 / 1160.27 / 0.64100 / 10035,044 / 33,297
Convection5k724103 / 1080.59 / 0.71100 / 1007,001 / 6,729
Allen–Cahn30k5,078112 / 1041.42 / 11.7495 / 4045,233 / 49,040
Allen–Cahn5k935108 / 1111.42 / 12.0195 / 558,648 / 8,397
Helmholtz30k4,656150 / 1586.37 / 9.4075 / 7531,073 / 29,539
Helmholtz5k85853 / 15816.23 / 25.6550 / 3016,222 / 5,437

Rebuttal Table R3-T3. Break-even = offline cost ÷ mean saving per query, over cases where both methods reach ≤5% error within 1,000 updates. One training seed; Helmholtz extrapolation rests on only 3 of 20 cases. Cohort sizes and cost details: Appendix C.2.1.

Against Standard-RBF, the 30k recipe breaks even after about 30,000–54,000 queries; stopping the same run at 5k reduces this about fivefold, with close accuracy on Burgers, Convection and Allen–Cahn. This comes mainly from the lower training cost, as the saving per query stays similar; only Helmholtz benefits from the longer training. A second training seed reproduces the 30k training cost within 1%.

In addition to this reporting, we also include the break-even points of the original 30k checkpoints against PINO and P²INN-SVD as a reference for the reader (Appendix C.2.1, Table 14b).

Overall, the main configuration does amortize, and its break-even point is mainly set by the training budget, which can be reduced while keeping close accuracy on three of the four equations.

The need for many queries is already stated as a limitation in the submitted paper. In the revised version, we update it to state that shorter training reduces the number of queries needed (Limitations, page 9, lines 479–481). Since this work focused on the initialization–refinement design rather than on optimizing the pre-training cost, we also mark this as a direction for future work; the shorter-training results above show that there is room to reduce it.

Full cost–accuracy results, including the 10k budget, cohort sizes and capped-cost sensitivity: Appendix C.2.1, Table 14a.

Reviewer_C9WxOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The authors propose a well motivated offline-online formulatio for solving multi-query parameterized PDEs, by leveraging radial basis functions. The authors train a task-conditioned generative initializer to predict the initial RBF parameters. During inference, the generated RBF state is refined using a physics-informed objective for a small number of steps.The method is evaluated on several PDE families, showing improved finite-budget adaptation compared to standard PINNs and neural operators.

Critical strengths

The paper is well written, and is largely easy to follow. The motivation is clear and the authors tackle an important problem.

The formulation allows for closed-form derivative computation instead of autograd operations, thus addressing one of the most important issues surrounding PINN stability.

The decoupled architecture is intuitive and the authors provide thorough experiments, ablations across a wide range of PDEs.

Critical weaknesses

While the generator inference is quite fast, the approach still requires online refinement. The model still requires expensive offline training. It would be interesting to see the overall amortization gains. How expensive would the joint offline-online setup be, in contrast to zero-shot models, and how much is the gain relative to it. Would a few-shot finetuning of a model like Walrus still underperform the current approach?

The choice of baselines seem a little unfair. For instance, the authors compare against FNO, and classical PINN. However, recently, there have been more expressive and powerful architectures. A more fair comparison could include some of these models:

  1. McCabe, Michael, Payel Mukhopadhyay, Tanya Marwah, Bruno Regaldo-Saint Blancard, Francois Rozet, Cristiana Diaconu, Lucas Meyer et al. "Walrus: A cross-domain foundation model for continuum dynamics." arXiv preprint arXiv:2511.15684 (2025).
  2. Herde, Maximilian, Bogdan Raonić, Tobias Rohner, Roger Käppeli, Roberto Molinaro, Emmanuel De Bezenac, and Siddhartha Mishra. "Poseidon: Efficient foundation models for pdes." Advances in Neural Information Processing Systems 37 (2024): 72525-72624.
  3. Koupaï, Armand Kassaï, Lise Le Boudec, and Patrick Gallinari. "Efficient Generative Transformer Operators For Million-Point PDEs." arXiv preprint arXiv:2512.04974 (2025).
  4. Viswanath, Hrishikesh, Hong Chul Nam, Xi Deng, Julius Berner, Anima Anandkumar, and Aniket Bera. "Operator learning using weak supervision from walk-on-spheres." arXiv preprint arXiv:2603.01193 (2026).

The zero-shot errors for GI-RBF before refinement are often quite high (e.g., 0.474 for Allen-Cahn, 0.587 for Burgers) compared to FNO. GI-RBF only becomes competitive after online refinement, but measuring an amortized fine-tuning method against pure zero-shot references makes evaluating the true state-of-the-art capability murky.

Questions for authors

Could you please provide more thorough amortization experiments, perhaps similar to what's done in "Operator learning using weak supervision from walk-on-spheres.".

Could you compare against more expressive and recent PINN and neural operator formulations?

2 comments

W1 and Q1 - Adaptation cost and amortization

Thank you for highlighting the combined offline and online cost. Our aim is to make per-query physics-based adaptation inexpensive: the generator runs once, while refinement updates only the compact RBF solver. This provides an alternative when zero-shot prediction does not meet the required accuracy, but the offline investment must be justified by repeated-query savings.

Following your suggestion, we added a time-to-accuracy comparison: we measure the complete query time needed to reach 5% relative-L2 error and report the fraction of cases reaching that target within the refinement budget. We compare GI-RBF with PINO and P²INN-SVD, which also perform physics-based adaptation. On Burgers and Convection, GI-RBF has the lowest median query time among successful cases, while target reach varies across methods (Appendix C, Table 12a).

As an additional amortization analysis, we include both methods’ offline costs and estimate when GI-RBF’s query-time savings recover its larger initial investment. We report these break-even estimates against PINO and P²INN-SVD where both methods reach the target (Appendix C.2.1, Tables 14a–14b).

Zero-shot prediction remains attractive when sufficiently accurate, as FNO is on Allen–Cahn; its training-cost records are unavailable for a break-even comparison. Fine-tuning Walrus is also an interesting alternative, but remains untested here. Our evidence establishes practical millisecond-scale refinement on Burgers and Convection and conditional amortization, rather than a universal advantage over zero-shot or foundation models.

W2 and Q2 - Baseline coverage

Thank you for the suggestions. Beyond FNO and the classical PINN, the submission already compares against recent physics-informed solvers and adaptive baselines: SPINN and PIG as solvers trained from scratch (Experiment I), and Hyper-LR-PINN, PINO and P²INN-SVD as learned baselines that adapt to new PDE instances (Experiment II). For our setting, these are the most direct baselines, because they are trained on the same data and adapted with the physics loss. We further strengthened this comparison with a matched time-to-accuracy evaluation of PINO and P²INN-SVD.

The suggested foundation models address a different setting. Walrus and Poseidon are pretrained on large and diverse datasets, so a comparison would change both the training data and the compute and would not isolate the effect of amortized initialization. WoS-NO trains a neural operator with physics-based Monte Carlo supervision for zero-shot prediction rather than adapting to each query. We now discuss the foundation models and WoS-NO in the related work (Section 2, page 2, lines 91–98.).

FNO and U-Net serve as accuracy and latency reference points: on Burgers and Convection, GI-RBF reaches 5% error within tens of milliseconds, within an order of magnitude of zero-shot inference, showing that inference through adaptation can remain practical.

Overall, the evaluated physics-informed adaptive baselines are the most direct comparison for amortized adaptation, while foundation models address transfer from large-scale pretraining.

W3 - Initial error and refinement

Our goal is an adaptive route to predicting solutions: a neural network initializes a compact numerical representation, which is then corrected using the target physics. GI-RBF's zero-shot output is therefore an initialization rather than a final prediction, and physics-based solving keeps a role at deployment, including when the initial prediction is insufficient.

We agree that this requires evaluating initialization and refinement together. We believe the zero-shot neural operator baselines (FNO and U-Net) serve as a line of reference for accuracy and latency, while the physics-informed baselines that are also fine-tuned at deployment (Hyper-LR-PINN, PINO and P²INN-SVD) provide the closer one-to-one comparison. The zero-shot references show where adaptation is useful: after refinement, GI-RBF has lower error than FNO on three of the four equations, while FNO is already more accurate on Allen–Cahn, where zero-shot prediction remains attractive.

We clarified in the captions of Tables 11–12 that FNO and U-Net are zero-shot predictors, while the other methods are refined with the physics loss (Table 11; Table 12).

We also add this to the limitations in the revised version: GI-RBF currently relies on refinement to reach competitive accuracy, and improving its zero-shot accuracy, which would further shorten refinement, is a direction for future work (Limitations, page 9, lines 481–483).

Thank you for the responses. I will maintain my accept score

Reviewer_BcZsOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes GI-RBF, which generates a task-conditioned initialization for an anisotropic RBF solver and then refines that initialization online using the PDE objective. The method is evaluated on Allen–Cahn, Burgers, Convection, and Helmholtz problems, including interpolation and extrapolation settings. The authors compare it with standard RBF initialization and several neural-network and neural-operator baselines, and report ablations, runtime measurements, offline training costs, and a Flow Mixing extension.

Critical strengths

The method has a clear motivation: task-conditioned initialization aims to reduce online optimization for new PDE instances while retaining physics-based refinement. The evaluation covers several PDE families and discusses ablations and offline costs, helping to clarify when the method may be useful.

Critical weaknesses

The main gap is between the backbone-selection motivation and the deployment experiments. RBF efficiency motivates the method, but the deployment experiments hold the RBF backbone fixed and do not compare different solver backbones under a matched amortized/online-refinement protocol. It therefore remains unclear how much of the claimed overall efficiency comes from choosing RBF, as opposed to the learned initialization alone.

Questions for authors

The paper presents RBF efficiency as a reason for its architectural choice. However, Figure 2 evaluates solvers at fixed update counts, while Experiment II holds the RBF backbone fixed. What evidence establishes that RBF is the more efficient backbone under the deployment setting where GI-RBF is claimed to be useful?

Tables 11–12 already include amortized non-RBF methods such as Hyper-LR-PINN, PINO, and P²INN-SVD, but their training, adaptation, and timing protocols differ. Can the authors provide a time-to-matched-accuracy comparison with at least two of these existing baselines under a more consistent deployment and timing protocol? If a controlled backbone ablation is infeasible, please clarify which efficiency conclusions concern GI-RBF as a complete system and which can be attributed specifically to the RBF backbone.

The break-even estimates are obtained from a reduced training configuration, whereas the main model has a substantially larger offline cost. What evidence supports an amortized efficiency claim for the main reported configuration?

3 comments

W1 and Q1 - Evidence for the RBF backbone

We chose RBFs as the backbone because their properties are useful for an inference-through-adaptation strategy. Different backbones could be used, for instance different kernel choices; however, it is useful for them to have a low parameter count, expressivity and good convergence properties. Different equations could benefit from different choices of solvers. The main goal of GI is to provide initial states that enable few-step, low-latency inference.

The solver comparison is already showcased in Table 2 with the improved solver. The standalone improved solver has a good floor, and GI helps reach the solution at a lower latency. In addition, it helps skip some of the difficult parts of the optimization, thus often reaching a better final accuracy.

To further support the distinction, we add a comparison with an MLP backbone under the same parameter count, using standard and GI initialization for both, with common test cases, online settings and training-update budget (Appendix D.5, Table 19c). Both the RBF backbone and learned initialization contribute to deployment efficiency.​

PDEModelMedian steps ↓Median query time (ms) ↓Target reached ↑
BurgersStandard RBF250 / 250124.8 / 125.265% / 67.5%
BurgersGI-RBF100 / 10045.1 / 45.672.5% / 70%
BurgersStandard MLP1,000 / 1,000831.9 / 823.77.5% / 12.5%
BurgersGI-MLP500 / 500413.8 / 406.560% / 52.5%
ConvectionStandard RBF500 / 250129.1 / 82.570% / 72.5%
ConvectionGI-RBF25 / 5012.5 / 17.895% / 100%
ConvectionStandard MLP500 / 500339.8 / 334.555% / 50%
ConvectionGI-MLP175 / 250110.3 / 157.0100% / 92.5%

Rebuttal Table R1-T1. First-recorded attainment of ≤5% error within 1,000 updates. Steps/time are medians among successful cases; reach includes all cases. Values are Interpolation / Extrapolation.

GI helps both solvers, and RBFs give lower median steps, lower latency and a higher 5% target reach in nearly every panel.

Q2 - Matched time-to-accuracy comparisons

Thanks for this suggestion. We agree that the time to a target error gives a better comparison. We extend the frozen checkpoints from `Tables 11–12* without retraining, using common test cases, physics-loss locations, loss weights, GPU and observation checkpoints. Each method retains its native optimizer; offline training is not matched:

PDEMethodMedian steps ↓Median query time (ms) ↓Target reached ↑
BurgersGI-RBF-256100 / 10052.4 / 52.790% / 92.5%
BurgersPINO100 / 250860.3 / 2,148.897.5% / 82.5%
BurgersP²INN-SVDNot reached / Not reachedNot reached / Not reached0% / 0%
ConvectionGI-RBF-25625 / 2517.4 / 18.1100% / 100%
ConvectionPINO100 / 250858.2 / 2,136.4100% / 100%
ConvectionP²INN-SVD10 / 25116.5 / 286.190% / 80%

Rebuttal Table R1-T2. First-recorded attainment of ≤5% error within 1,000 updates. Steps/time are medians among successful cases; reach includes all cases. Values are Interpolation / Extrapolation.

On Burgers and Convection, GI-RBF reaches the target with the lowest query time in every panel. The appendix extends this comparison to Allen–Cahn and Helmholtz: on Helmholtz, GI-RBF is the only method that reaches the target, while on Allen–Cahn PINO reaches it more often.

Overall, GI-RBF is the fastest method to reach the target accuracy on three of the four equations, supporting low-latency inference through adaptation.

Full results for all four equations and timing boundaries: Appendix C, Table 12a.

Q3 - Amortization of the main configuration

The initial 30k recipe used a fixed training budget. This choice was rather conservative, which inflated the break-even point. The reduced 3k training was introduced to show that the break-even point is a direct function of the GI training, which can be further optimized and thus result in a more attractive break-even point. We therefore present transparent reporting of the break-even point and accuracy at the higher-end (30k) and lower-end (5k) budgets of the same recipe, where the only modification is the number of training updates.

PDETraining updatesOffline cost (s)Saving/query vs Standard-RBF (ms)Error @1,000 (%)Reach ≤5% (%)Break-even queries
Burgers30k5,640188 / 1052.60 / 1.7685 / 10029,933 / 53,775
Burgers5k1,018181 / 932.82 / 1.8185 / 955,627 / 10,996
Convection30k3,863110 / 1160.27 / 0.64100 / 10035,044 / 33,297
Convection5k724103 / 1080.59 / 0.71100 / 1007,001 / 6,729
Allen–Cahn30k5,078112 / 1041.42 / 11.7495 / 4045,233 / 49,040
Allen–Cahn5k935108 / 1111.42 / 12.0195 / 558,648 / 8,397
Helmholtz30k4,656150 / 1586.37 / 9.4075 / 7531,073 / 29,539
Helmholtz5k85853 / 15816.23 / 25.6550 / 3016,222 / 5,437

Rebuttal Table R1-T3. Each row uses its own fixed configuration and costs. Values are Interpolation / Extrapolation. Break-even = offline cost ÷ mean saving per query, over cases where both methods reach ≤5% error within 1,000 updates. One training seed; Helmholtz extrapolation rests on only 3 of 20 cases. Cohort sizes and cost details: Appendix C.2.1.

Against Standard-RBF, the 30k recipe breaks even after about 30,000–54,000 queries; stopping the same run at 5k reduces this about fivefold, with close accuracy on Burgers, Convection and Allen–Cahn. This comes mainly from the lower training cost, as the saving per query stays similar; only Helmholtz benefits from the longer training. A second training seed reproduces the 30k training cost within 1%.

In addition to this reporting, we also include the break-even points of the original 30k checkpoints against PINO and P²INN-SVD as a reference for the reader (Appendix C.2.1, Table 14a).

Overall, the main configuration does amortize, and its break-even point is mainly set by the training budget, which can be reduced while keeping close accuracy on three of the four equations.

The need for many queries is already stated as a limitation in the submitted paper. In the revised version, we update it to state that shorter training reduces the number of queries needed (Limitations, page 9, lines 479–481). Since this work focused on the initialization–refinement design rather than on optimizing the pre-training cost, we also mark this as a direction for future work; the shorter-training results above show that there is room to reduce it.

Full cost–accuracy results, including the 10k budget, cohort sizes and capped-cost sensitivity: Appendix C.2.1, Table 14a.

Your answer resolved my question, and I have updated the score. In addition, I have one more suggestion: the presentation of the flowchart could be optimized.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.