acceptodds
Under review as a conference paper at ICLR 2027

Spectral Allocation and Routing for Physics-Informed Learning

Abstract

Physics-informed neural networks (PINNs) often struggle with spatially varying high-frequency solutions of partial differential equations (PDEs) due to spectral bias. Existing approaches, such as Fourier features, periodic activations, multiscale architectures, and adaptive Fourier representations, mitigate this bottleneck but often rely on either globally learned spectral representations or locally adapting predefined spectral bases. We introduce Spectral Allocation and Routing (SAR), an adaptive mechanism that learns a global set of Fourier frequencies and controls their local contributions through continuous, frequency-conditioned spatial routing. This formulation enables direct inspection of the learned spectral content and provides a straightforward way to incorporate spectral inductive biases through structured support parameterizations. Controlled experiments on nonstationary analytical targets investigate spectral discovery and local routing, while neural tangent kernel (NTK) analysis examines their effects on learning dynamics. We further evaluate SAR on heterogeneous Helmholtz problems, where it substantially improves reconstruction accuracy over unrouted representations and achieves low errors across spatially varying and localized multiscale wavefields.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
Physics-informed neural networks, Spectral bias, Adaptive spectral learning, Spectral routing, Fourier representations, Partial differential equations, Scientific machine learning

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_2nssOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Spectral Allocation and Routing (SAR) for physics-informed neural networks (PINNs), an adaptive mechanism that learns a global set of Fourier frequencies and controls their local contributions through continuous, frequency-conditioned spatial routing. Common Fourier feature embedding methods use prior bandwidth to fix the spectrum of the network before training, the network cannot adjust this bandwidth during training, which may lead to wasted high-frequency features. In contrast, SAR jointly learns trainable frequencies and the weights assigned to them. The paper also includes an analysis of the network’s neural tangent kernel (NTK). Together, these are intended to demonstrate that the proposed method mitigates the spectral bias of the network.

Critical strengths
  1. The proposed SAR introduces a joint formulation that couples globally trainable Fourier frequencies with continuous spatial-frequency routing, which enhances the effectiveness of Fourier feature embedding.
  2. The paper introduces optional geometric supports that organize groups of frequencies within a shared spectral structure.
  3. The paper uses NTK analysis to examine the architecture of the proposed method, which provides some interpretability regarding the mitigation of spectral bias.
Critical weaknesses
  1. Unlike the analysis given in Tancik et al. (2020), the NTK analysis in this paper does not examine the eigenvalue decay of the network’s NTK. Instead, the analysis is based only on relatively simple one-dimensional linear regression problem, which does not explain how SAR improves the NTK of the MLP during training process.
  2. The presentation of the method is not very clear. The content of Section 3.2 is not well supported by the background material presented earlier in the paper.
  3. The benchmarks and baselines are limited in this paper, with only two-dimensional heterogeneous Helmholtz problems under manufactured solutions. No other popular PINN benchmarks are considered (there are a lot popular benchmarks regarding the spectral bias and high-frequency problem). The choice of baselines and the comparison also limit the conclusions.
Questions for authors
  1. In Section 3.2, could the authors explain why the RBF formulation in frequency space was adopted for the Helmholtz problems instead of the routers studied in Section 4.2? More generally, how should one choose among these router designs and their hyperparameters for more experiments?
  2. Regarding Weakness 3, could the authors evaluate SAR and the baselines on a broader set of problems? It could include many commonly used PINN benchmarks. Including similar recent methods identified in the related work.
  3. The problem setting in Section 3.1 considers only spatial coordinates and steady-state equations. Can SAR be applied to time-dependent PDEs? Would time be treated as an additional input dimension with space-time wavevectors and space-time routing, or would temporal frequencies be handled separately?

Overall, the SAR method seems to be an extension based on the Fourier feature embedding. Therefore, it's good to see the authors can provide more general applicability of SAR similar to the Fourier feature methods.

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

This work addresses spectral bias for PINNs by proposing Spectral Allocation and Routing (SAR), which jointly optimizes global Fourier wavevectors and a continuous frequency‑conditioned routing module to modulate frequency contributions spatially. Comprehensive experiments on Helmholtz problems are conducted, including ablation study and NTK analysis.

Critical strengths
  1. This paper proposes a novel method to replace random Fourier embeddings, improving performance and alleviating spectral bias.

  2. The SAR method adapts during training (especially K\mathcal{K}), which is a novel approach to modeling training in PINNs.

  3. A comprehensive ablation study has been conducted.

Critical weaknesses
  1. This paper is a little difficult to follow, as some symbols and parameters are not clearly defined. For example, it is unclear which parameters in Equation (3) are learnable and what the operator ⊘\oslash means?

  2. The experiments are conducted only on steady-state equations. The performance of the proposed method on time-dependent equations remains unclear.

  3. The routing function rψr_{\psi} is specifically designed for Helmholtz equations. However, the generalizability of this routing function to other types of PDEs is unclear, and the authors do not provide a clear justification for choosing this particular function. Especially, the authors do not sufficiently compare against other radial basis functions (e.g., multiquadric, inverse multiquadric).

Questions for authors
  1. How would the SAR architecture need to be modified to handle time-dependent PDEs? Will K\mathcal{K} depend on tt?

  2. Can SAR still achieve impressive performance on other types of PDEs when using the Gaussian radial basis functions defined in Equation (4)?

Reviewer_Bt3NOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Spectral Allocation and Routing (SAR), which mitigates spectral bias in PINNs by jointly learning a global set of Fourier frequencies and a frequency-conditioned spatial router, thereby deciding both which frequencies to represent and where to use them. Controlled 1D experiments and a projected NTK analysis separately demonstrate global frequency discovery by structured supports and local frequency tracking by the router. On 2D heterogeneous Helmholtz benchmarks, routing substantially reduces error relative to unrouted variants and outperforms frequency-blind baselines, although carrier-informed Gabor baselines remain competitive on some problems.

Critical strengths
  1. The formulation cleanly separates frequency allocation from spatial routing while training them jointly, and the free/geometric parameterizations provide a natural way to inject spectral priors.
  2. The controlled analytical experiments isolate the global-discovery and local-allocation mechanisms, and the NTK analysis is presented with its interpretive limitations clearly stated.
Critical weaknesses
  1. The benchmarks are limited to two types of 2D Helmholtz problems with manufactured solutions, so claims of generality to physical scattering problems, other PDE families, or higher-dimensional and time-dependent settings are not supported.
  2. The unrouted variants and several external baselines exhibit very large errors, making it unclear whether they were adequately optimized under the common training protocol; in addition, as the authors note for E1B, routed and unrouted models differ in their initial feature scaling, which confounds that comparison.
  3. There is no direct comparison with the closest prior methods discussed in the related work that address both global frequency adaptation and spatial adaptation, and the included baselines are substantially modified from their original implementations, which limits the persuasiveness of the comparison.
  4. The source of the routing gain is not isolated: there is no ablation that removes the spatial or frequency dependence of the router separately, the NTK analysis is not extended to the routed Helmholtz models, and no spatial maps of routing weights are shown for the Helmholtz problems to verify that local routing actually occurs.
  5. There is no compute-matched baseline accounting for the additional cost of routing, some controlled experiments are reported with a single seed, and the reason why the relative ranking of geometric and free parameterizations differs across experiments is not examined.
Questions for authors
  1. Were the learning rate or initialization scale tuned separately for the unrouted and external baselines, and in E1B, does the advantage of routing persist when the initial feature scaling is matched to the unrouted models?
  2. If the router is ablated to depend only on position or only on frequency, which component accounts for most of the gain, and can spatial maps of the routing weights on the particle problems confirm that high-frequency modes are activated only in localized regions?
  3. Can the authors add a direct comparison with the closest prior methods and a compute-matched baseline (e.g., more modes or more training steps)?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.