acceptodds
Under review as a conference paper at ICLR 2027

KAPI-ONet: Kernel-Adaptive Physics-Informed Operator Network for PDEs

Abstract

Physics-informed operator learning enables rapid prediction across families of PDEs without requiring paired numerical solution fields, but existing methods often rely on latent or globally shared representations that are difficult to interpret physically and can be computationally expensive to train. This raises the question of whether an operator learner can expose how variations in the input reorganize the solution representation while remaining compact enough to reduce the computational burden of physics-informed training. To address this gap, we propose KAPI-ONet, a kernel-adaptive physics-informed operator network in which each solution is represented explicitly by a task-dependent collection of Gaussian RBFs with interpretable centers, scales, orientations, and coefficients in physical or space–time coordinates. For functional inputs, KAPI-ONet formulates operator learning as a mapping between structured Gaussian-splat representations of the input and solution, providing a common coordinate-aware description of both sides of the operator through geometrically meaningful amplitudes, locations, and scales. Across elliptic, time-dependent, and Navier–Stokes problems, the learned kernel geometry adapts to localized and high-gradient solution structures, while KAPI-ONet attains competitive accuracy with substantially lighter training in settings where direct computational comparisons are available. We also evaluate KAPI-ONet beyond standard forward PDE operator learning, including parameter-to-field prediction and two biomedical applications: inverse hemodynamics from sparse velocity measurements and weakly supervised cardiac annotation. These results support task-adaptive kernel geometry as an interpretable and computationally efficient representation for operator learning across diverse problem formulations. Code and reproducibility materials are available at: https://anonymous.4open.science/status/kapi-onet-anonymous-6C97

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
physics-informed operator learning, radial basis functions, Gaussian representations, parametric PDEs, interpretable neural operators

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_BQrkOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes KAPI-ONet, a physics-informed operator-learning framework that directly generates an explicit, task-adaptive Gaussian radial-basis representation of the solution field. For functional inputs, a pretrained Gaussian-splat functional autoencoder (GSFAE) encodes inputs into interpretable amplitudes, centers, and scales. A conditioning network then predicts a compact “parent–child” kernel geometry and coefficients that decode into an analytic continuous solution field, trained only with PDE and boundary/initial residuals. Experiments across elliptic, time-dependent, and Navier–Stokes problems, plus two biomedical applications (inverse hemodynamics and weakly supervised cardiac annotation), show that KAPI-ONet yields competitive accuracy with substantially reduced CPU-only training time and provides interpretable task-dependent kernel geometry.

Critical strengths

The core idea of task-conditioned, explicit kernel geometry for operator learning is well motivated and distinct from dominant latent-representation approaches (e.g., DeepONet trunk, FNO transforms). The hierarchical parent–child factorization is a thoughtful design to balance geometric adaptability with parameter efficiency, offering a principled mechanism to share geometry while keeping child coefficients independent. The mapping from a structured Gaussian-splat input representation to a similarly structured solution representation yields an interpretable, coordinate-aware interface on both sides of the operator.

Critical weaknesses
  1. Freezing the GSFAE may create a representational bottleneck, particularly OOD where previously unseen fine scales appear; the paper acknowledges degradation on reduced correlation-length GRFs.
  2. The approach is demonstrated mainly on 1D/2D and relatively smooth problems; how the fixed Gaussian family handles sharp discontinuities, complex 3D geometries, or multi-physics couplings is unclear to me.
  3. There is no theoretical guarantees (e.g., approximation rates or stability with parent–child sharing) provided. Selection of the number of parents/children is empirical.
  4. Baseline comparisons are not fully apples-to-apples. PI-WNO uses a different error metric and some hardware/training-budget normalizations are estimated or missing. Key modern PINO baselines (e.g., PINO, PI-CViT) are not reported.
  5. There is no ablation of the GSFAE itself (e.g., joint end-to-end training vs freezing; reconstruction fidelity vs operator accuracy), nor a study of collocation strategies and loss weighting (which strongly affect PI training).
  6. Comparisons against RBF/PINN hybrids (SPINN, sparse RBF PINNs) or kernel-operator baselines are absent despite high conceptual overlap.
  7. Biomedical experiments lack robustness studies to measurement noise and broader clinical validation. The reported hemodynamics involve idealized settings and sparse/noiseless measurements.
Questions for authors
  1. What are the typical parent/child kernel counts and parameter budgets used per benchmark, and how sensitive are results to these choices? Could you report accuracy vs kernel count trade-offs in the main text?
  2. How accurate is the GSFAE reconstruction on the input distributions used, and how does end-to-end fine-tuning (vs freezing) affect operator accuracy and OOD robustness?
  3. Can you clarify the exact decoding formula (A_z, B_z, child sum) and which boundary/initial conditions are hard-embedded versus softly penalized in each benchmark?
  4. Could you provide unified metric comparisons against PINO/PI-CViT (and, if feasible, SPINN/sparse RBF PINNs) under matched training budgets, even at reduced scales, to better situate KAPI-ONet’s accuracy/efficiency trade-off?
  5. How are collocation points sampled and weighted across interior/BC/IC residuals, and how sensitive is training to these choices (e.g., Latin hypercube vs uniform, adaptive sampling)?
  6. For hemodynamics, how robust is WSS recovery to measurement noise and sensor misplacement? Can you include a small parametric study (e.g., SNR sweep, sensor proximity to walls)?
  7. Do you foresee a principled criterion for selecting the degree of parent–child sharing (e.g., validation-based model selection or information criteria), and could this be automated during training?
16 comments
AuthorsOpenReview ↗

Summary of first response to Rev BQrk

We sincerely thank the reviewer for the constructive assessment. We have substantially strengthened the manuscript through additional experiments, theoretical analysis, and methodological clarifications. We expanded the parent--child factorization study across three PDEs (Appendix H), quantified GSFAE reconstruction accuracy, and compared frozen and jointly physics-trained representations, including their OOD performance (Appendices E and G). We also clarified hierarchical decoding, boundary and initial-condition enforcement, and physics-informed training protocols (Appendix C), and added hemodynamic sensitivity studies involving measurement noise and sensor placement (Appendix E). New theoretical results address approximation and stability (Appendix I), while Appendix J expands the methodological comparisons and investigates adaptive basis and collocation refinement. To ensure metric-consistent benchmarking, we use published baseline results to preserve their optimized configurations and avoid under-tuned reproductions. We re-evaluated KAPI-ONet using the corresponding published error metrics and clearly distinguish measured from projected computational costs. MAIN CHANGES IN MANUSCRIPT: BLUE COLOR.

AuthorsOpenReview ↗

Response to Q1 of Rev BQrk

We thank the reviewer for this suggestion. We have clarified the parent–child kernel counts and trainable parameter budgets for all benchmarks in Appendix F, particularly Tables F.2 and F.3. The forward PDE benchmarks typically use 2525 parents with 88 children each (200200 physical RBFs), whereas the cavity-flow and inverse-hemodynamics benchmarks use 5050 parents with 88 children each (400400 RBFs).

To assess sensitivity to these choices, we have expanded the structural ablation from Burgers to three PDE benchmarks: Burgers, diffusion--reaction, and Poisson. We consider five parent--child factorizations while keeping the total number of physical RBFs fixed at 200200. The results reveal a problem-dependent accuracy--parameter trade-off: Burgers benefits progressively from additional parents, while diffusion--reaction exhibits relatively modest, non-monotonic variations and Poisson largely saturates around 2525 parents.

We have added a quantitative summary to the main text (Section 3.1, paragraph Kernel factorization and parameter budget). The complete numerical results and accuracy--parameter-budget curves are provided in Appendix H, Table H.1 and Fig. H.1. These experiments are intended to characterize the trade-off between geometric flexibility and parameter efficiency, rather than identify an operating point optimized for the lowest test error.

AuthorsOpenReview ↗

Response to Q2 of Rev BQrk

We thank the reviewer for this important question. We have quantified GSFAE reconstruction accuracy across the benchmark input distributions. The selected reconstruction-pretrained representations achieve mean held-out relative errors of 0.17%0.17\% percent for parametric Burgers, 1.63%1.63\% percent for GRF Burgers, 1.34%1.34\% percent for diffusion--reaction, and 5.56%5.56\% percent and 5.61%5.61\% percent for localized and sinusoidal Poisson, respectively. The capacity studies and reconstruction results are provided in Appendix E, with architecture details in Appendix F.

To assess the effect of jointly learning the input representation, we compare reconstruction-pretrained, frozen GSFAE with joint physics-only GSFAE--KAPI-ONet training across four PDE benchmarks. Appendix G now includes both a quantitative comparison (Table G.1) and four new qualitative figures showing identical input functions, reference solutions, predictions from both models, and their absolute-error fields.

The quantitative results demonstrate substantial improvements under challenging distribution shifts. For combined OOD shifts, mean relative L2L^2 error decreases from 72.45%72.45\% percent to 58.14%58.14\% percent for periodic Burgers, from 34.78%34.78\% percent to 17.94%17.94\% percent for diffusion--reaction, and from 40.46%40.46\% percent to 24.48%24.48\% percent for Poisson. The new paired figures visualize these improvements in the predicted solution structures and error distributions. They also illustrate the problem-dependent behavior observed for parametric Burgers, where joint training improves some individual extrapolation cases but yields similar aggregate errors.

Importantly, the joint experiments optimize GSFAE and KAPI-ONet together from scratch using only the physics-informed objective, rather than fine-tuning a reconstruction-pretrained encoder. The results indicate that physics-guided input representations can improve downstream operator accuracy and OOD generalization, even when accurate reconstruction of the original input function is not their optimization objective. Broader comparisons, including fine-tuning pretrained encoders, remain a promising direction for future work.

Reviewer_DfEhOpenReview ↗
Rating
4
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This is a strong and interesting paper proposing KAPI-ONet, an operator-learning framework that adapts the location, scale and orientation of Gaussian basis functions to the structure of each PDE problem. The resulting representation is both interpretable and flexible, and the authors demonstrate it across a broad range of experiments, including pure PINN problems as well as supervised training on cardiac Magnetic Resonance imaging.

Critical strengths

The paper is well written and well structured, brings a novel approach and method to the community, and will be of great interest to the ICLR audience. It is demonstrated on a large variety of use cases, demonstrating not only generality but also significant efforts from the authors to apply their method to a range of applications. The writing style is engaging, it was a great read.

Critical weaknesses

No critical weaknesses, just some curious comments below.

Questions for authors

Given that KAPI-ONet learns to concentrate its representation around localised solution structures, it would be interesting to explore how this advantage compares to simple adaptive spatial sampling, for example with the localised Poisson problems. This is not required for the paper, unless the authors feel they can achieve this rapidly, but it’s an interesting question. In fact, it may be interesting to understand whether adaptive sampling and kernel adaptation could provide complementary or combined benefits in future works.

1 comment
AuthorsOpenReview ↗

Response to Rev DfEh

We sincerely thank the reviewer for the encouraging assessment of our work and the thoughtful suggestion. In response, we have added a preliminary numerical study in Appendix J.1 (Figs. J.1 and J.2) illustrating how adaptive Gaussian-kernel geometry and gradient-guided collocation can work together to capture localized moving structures. Using periodic advection as a representative example, we demonstrate their compatibility while maintaining fixed kernel and collocation budgets, achieving approximately 0.59%0.59\% relative L2L^2 error. Although a systematic comparison of the individual and combined benefits remains for future work, this experiment provides an initial step toward integrating adaptive sampling with KAPI-ONet's task-adaptive representation. MAIN CHANGES IN MANUSCRIPT: MAGENTA COLOR.

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

KAPI-ONet generates the centers, scales, orientations, and coefficients of Gaussian radial basis functions from task inputs, representing each output as an explicit continuous kernel field. For functional inputs, a Gaussian-splat functional autoencoder (GSFAE) is pretrained and frozen before a conditioner generates the solution state. A parent–child construction shares geometric transformations to reduce the number of generated geometric variables. Training uses PDE residuals or variational objectives. Experiments cover Poisson, diffusion–reaction, Burgers, parametric Navier–Stokes, sparse-observation inverse hemodynamics, and weakly supervised cardiac annotation. The results demonstrate the potential of a compact representation, relatively inexpensive CPU training, and the use of a common output representation across different tasks.

Critical strengths
  1. The continuous kernel representation is explicit, and the parent–child construction has a clear compactness benefit. The conditioner generates spatial or space–time kernel geometry that can be inspected directly, and the resulting solution state can be reused across query locations. For a scalar output with N=200N=200 physical kernels and K=25K=25 parents, the generated state decreases from 6N+1=12016N+1=1201 quantities for independently parameterized kernels to N+5K+1=326N+5K+1=326. Appendix H varies geometric sharing at a fixed physical kernel budget and demonstrates an empirical trade-off between parameter count and error.

  2. The implementation covers diverse tasks and provides useful feasibility results. The forward PDE experiments do not use paired solution-field supervision. Sinusoidal Poisson and the sample-count-matched diffusion–reaction experiment achieve mean relative L2L^2 errors of 0.39% and 0.76%, with complete CPU training pipelines of approximately 45.8 minutes and under 40 minutes, respectively. Cardiac annotation uses a subject-level split: on 75 images from 15 independently held-out subjects, the frozen model achieves a mean Dice of 0.868, increasing to 0.912 after additional per-image variational correction. These results support the representation’s application potential across different tasks.

Critical weaknesses
  1. The independent contribution of the central adaptive-geometry design is insufficiently established. All variants in Appendix H retain task-dependent geometry, and changing the parent count also changes the number of network parameters. This ablation therefore establishes a sharing trade-off but does not determine whether globally learned, task-shared kernel geometry with only task-dependent coefficients would suffice. Similarly, GSFAE lacks controlled comparisons against direct sensor inputs or alternative encodings of comparable dimension. The current results cannot adequately separate the contributions of task-dependent geometry, input encoding, and the Gaussian output representation. This directly affects the assessment of the methodological advance and calls for targeted controls with the same physical kernel budget and comparable optimization conditions.

  2. The efficiency comparisons do not yet determine the computational savings under a common accuracy requirement. Appendix E.2 reports approximately 40 hours for a complete PI-DeepONet run on the same CPU, versus under 40 minutes for KAPI-ONet including GSFAE capacity selection and pretraining. This is meaningful evidence of a cost difference between the compared training schedules. However, the reported diffusion–reaction errors are 0.45% and 0.76%, respectively, and it is unclear how long the baseline would need to reach 0.76% if stopped earlier. Sinusoidal Poisson is compared using a common metric, but the corresponding errors are 0.0469% and 0.366%, and the PI-WNO total runtime is projected from per-epoch measurements. The parametric Burgers comparison also differs in initial-condition construction and sample counts. These endpoint comparisons demonstrate a useful cost–accuracy trade-off, but do not establish whether the relative efficiency advantage persists under a common accuracy requirement. This uncertainty affects the assessment of the size of the central efficiency contribution.

Questions for authors
  1. With the same GSFAE, physical kernel count, constraint handling, and training budget, how do error and runtime change if output geometry is globally learned and shared across tasks, while only coefficients remain task-dependent? Please also report parameter counts and differences across repeated training runs to assess the benefit of adaptive geometry itself.

  2. Does GSFAE improve solution accuracy or total training cost relative to direct sensor inputs or a conventional encoder with a comparable representation dimension? Appendix G already shows substantial input-reconstruction degradation for rougher out-of-distribution inputs. Can controlled comparisons distinguish the input-representation bottleneck from errors in the learned output operator?

  3. For at least diffusion–reaction or sinusoidal Poisson, can you provide error-versus-measured-training-time curves using the same task distribution, test samples, metric, and hardware for both methods? Using your existing accounting that includes GSFAE costs, please compare time to a common accuracy target to complement the endpoint comparisons of the full training schedules.

7 comments
AuthorsOpenReview ↗

Summary of first response to Rev WYQe

We thank the reviewer for the constructive comments, which motivated substantial new experiments presented in Appendix K. First, controlled parameter-matched ablations demonstrate that task-adaptive Gaussian-RBF geometry reduces the mean relative L2L^2 error from 6.784%6.784\% percent to 3.439%3.439\% percent compared with globally shared geometry, with the improvement confirmed across three independent training seeds (3.515±0.101%3.515\pm0.101\% percent versus 6.887±0.191%6.887\pm0.191\%percent) . Second, input-representation ablations show that GSFAE outperforms a conventional MLP encoder of comparable capacity (3.439%3.439\% percent versus 3.999%3.999\%percent) , although direct sensor conditioning achieves a lower error (2.993%2.993\%percent) , demonstrating that GSFAE is useful but not universally optimal. An additional diffusion--reaction study quantifies how input-reconstruction errors propagate to the numerical PDE solution, helping distinguish representation sensitivity from downstream operator approximation. Third, we provide a detailed physics-residual convergence analysis and computational-cost accounting, including GSFAE pretraining. KAPI-ONet reaches a training residual threshold of 10−410^{-4} in approximately 26,80026{,}800 updates, compared with 44,10044{,}100 estimated from the published PI-DeepONet trajectory, with an estimated KAPI-ONet pipeline time of 97.1197.11 minutes versus 15.4515.45 projected hours for PI-DeepONet. We explicitly distinguish measured, allocated, and projected times and acknowledge that these physics-based convergence comparisons do not establish a definitive speedup at matched held-out solution accuracy. Collectively, these additions strengthen the empirical evidence for task-adaptive geometry, clarify the benefits and limitations of GSFAE, and provide a more transparent assessment of KAPI-ONet's accuracy--efficiency trade-off. MAIN CHANGES IN MANUSCRIPT: RED COLOR.

AuthorsOpenReview ↗

Response to Q1 of Rev WYQe

We thank the reviewer for this important suggestion. We have added controlled architectural ablations in Appendix K.2 (Table K.1 and Fig. K.1) to isolate the contribution of task-adaptive output geometry. All variants retain the same frozen GSFAE, N=200N=200 physical Gaussian RBFs, constraint handling, physics-informed training objective, and 20,00020{,}000 optimization updates. In the globally shared variant, parent centers, scales, and orientations are learned jointly across tasks but remain task-independent, while all RBF expansion coefficients remain task-dependent.

In the original comparison, globally shared geometry achieves a mean relative L2L^2 error of 7.148 percent7.148\text{ percent} with 89,42989{,}429 trainable parameters, compared with 3.439 percent3.439\text{ percent} and 108,179108{,}179 parameters for task-adaptive KAPI-ONet. To eliminate the parameter-count confound, we additionally construct a parameter-matched shared-geometry model with 108,149108{,}149 parameters, only 3030 fewer than KAPI-ONet. Under this controlled comparison, task-adaptive geometry reduces the mean error from 6.784 percent6.784\text{ percent} to 3.439 percent3.439\text{ percent}, an improvement of 49.3 percent49.3\text{ percent}. The corresponding measured CPU operator-training times are 71.6871.68 and 90.4590.45 minutes, respectively, reflecting the additional computational cost of task-dependent geometry.

To assess reproducibility, we repeated the parameter-matched comparison across three independent training seeds (Appendix K.3, Table K.2). Task-adaptive geometry consistently outperforms globally shared geometry, achieving mean relative L2L^2 errors of 3.515±0.101 percent3.515\pm0.101\text{ percent} versus 6.887±0.191 percent6.887\pm0.191\text{ percent} (mean ±\pm sample standard deviation). The mean paired improvement is 3.372±0.2473.372\pm0.247 percentage points, with all three seeds favoring adaptive geometry. These results provide direct evidence that the accuracy advantage arises from task-dependent kernel geometry rather than simply from a larger trainable parameter budget.

AuthorsOpenReview ↗

Response to Q2 of Rev WYQe

We thank the reviewer for this important question. We have added controlled input-representation ablations in Appendix K.4 (Table K.3). On periodic Burgers, direct sensor conditioning achieves 2.993%2.993\% mean relative L2L^2 error, compared with 3.439%3.439\% for frozen GSFAE. However, GSFAE outperforms a conventional MLP encoder with matched latent dimension and encoder parameter count (3.439%3.439\% versus 3.999%3.999\%), while achieving better input reconstruction (1.629%1.629\% versus 2.466%2.466\%). These comparisons clarify GSFAE's representational benefits and limitations. Direct conditioning also avoids encoder pretraining, so GSFAE's advantages concern structured representation rather than guaranteed computational savings.

To further distinguish input-representation effects from operator-prediction errors, Appendix K.5 (Table K.4) introduces a numerical error-propagation study on 120120 held-out diffusion--reaction tasks. The mean GSFAE reconstruction error is 0.724%0.724\%, while solving the PDE with reconstructed inputs produces a 0.556%0.556\% mean solution perturbation, compared with 0.516%0.516\% KAPI-ONet prediction error. The strong correlation between input reconstruction error and induced solution perturbation (Pearson r=0.899r=0.899) confirms the physical significance of representation fidelity. This controlled in-distribution diagnostic separates reconstruction-induced PDE sensitivity from observed prediction error, while recognizing that representation and operator errors may interact. Together with the OOD experiments in Appendix G, these results provide a more detailed characterization of GSFAE's contribution to operator learning.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.