acceptodds
Under review as a conference paper at ICLR 2027

IGA-ODIL: Fast Physics-Informed Learning with B-Spline Representations

Abstract

Physics-informed neural networks (PINNs) formulate the solution of partial differential equations as residual minimization problems over neural network parameterizations. Although highly flexible, optimization of PINNs using modern variants of Stochastic Gradient Descent algorithms is expensive. On the other hand, iterative computation of PINN parameterization using the Gauss-Newton method suffers from convergence difficulties, dense Jacobian structures, and poor conditioning that limit the effectiveness of second-order optimization methods. In this work, we introduce IGA-ODIL (Optimization of DIscrete Loss with data representation following from IsoGeometric Analysis). Instead of neural-network parameterizations of PINNs, the unknown solution is represented by smooth B-spline basis functions, leading to structured, sparse Jacobians and efficient second-order Gauss–Newton optimization. For the nonlinear inverse problem, we use a damped Gauss–Newton (Levenberg–Marquardt) variant to improve robustness. In the considered implementations, the resulting sparse second-order optimization achieves two to three orders of magnitude lower wall-clock time than the corresponding first-order PINN and CRVPINN implementations. The resulting systems inherit locality, sparsity, and approximation-theoretic properties of classical finite element and isogeometric methods while preserving the residual-learning philosophy of scientific machine learning. The proposed methodology is evaluated on several benchmark problems, including Poisson equations, convection-dominated advection–diffusion equations, Helmholtz problems with highly oscillatory solutions, nonlinear Allen–Cahn equations, and inverse Helmholtz parameter identification. Numerical experiments demonstrate orders-of-magnitude speedups compared with PINNs and CRVPINNs while maintaining high accuracy and robustness.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
Physics Informed Neural Networks, Robust loss, Gauss-Newton method, B-splines, Isogeometric analysis, Residual learning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_VPTtOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes IGA-ODIL to address the optimization difficulties of PINNs, particularly the dense and poorly conditioned Jacobians induced by neural-network parameterizations that limit the effectiveness of second-order optimization methods. IGA-ODIL replaces neural-network parameterizations with B-spline representations and optimizes the spline coefficients using Gauss–Newton methods. IGA-ODIL is evaluated on a lot of benchmarks. However, the paper suffers from extremely poor presentation and significantly lacks sufficiently comprehensive comparisons. Therefore, the reviewer believes that the paper requires substantial revision and is not yet ready for acceptance at ICLR.

Critical strengths
  1. The limitations described in this paper correspond to real problems in existing physics-informed learning methods.
  2. The motivation of IGA-ODIL is reasonable.
Critical weaknesses
  1. The paper suffers from extremely poor presentation and organization. First, the Introduction contains a large amount of discussion that is not directly relevant to the core argument, making the main logic of the paper difficult to follow. The reviewer suggests simplifying the Introduction to focus on the central motivation and moving broader discussions of existing approaches to a dedicated Related Work section. In addition, Section 2 largely repeats descriptions of existing methods while occupying a substantial portion of the paper. In contrast, the most important part (the description of IGA-ODIL) is overly compressed. The experimental section is also poorly organized and lacks ablation studies on the core components. Therefore, the reviewer believes that the paper requires substantial reorganization.
  2. IGA-ODIL appears to be interpretable as a special type of explicit basis-function method. In particular, IGA-ODIL represents the solution using a set of B-spline basis functions, while the optimization mainly operates on the spline coefficients. Based on this formulation, many alternative basis functions could potentially be used to construct similar methods, including RBFs, Fourier bases, and wavelets. The reviewer would like the authors to further analyze why B-splines are selected and what advantages they provide compared with other basis functions.
  3. The paper lacks sufficiently comprehensive baseline comparisons, making the experimental evaluation highly insufficient and preventing an effective assessment of whether IGA-ODIL truly outperforms existing methods.
Questions for authors

Please refer to the weaknesses above.

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

This paper proposes IGA-ODIL, a physics-informed PDE solver that replaces neural network representations with B-splines. By exploiting the local support of B-spline basis functions, the method constructs sparse Jacobians and uses Gauss Newton optimization to solve the resulting residual minimization problem. The framework is also extended to nonlinear and inverse problems. Experiments demonstrate speedups over PINNs and CRVPINNs while maintaining good accuracy.

Critical strengths
  1. The paper addresses the high computational cost of physics-informed learning by replacing neural network optimization with a more efficient spline-based approach.

  2. The use of B-spline representations leads to sparse Jacobians and enables efficient Gauss Newton optimization. This is a practical approach that takes advantage of the smoothness and local support of spline basis functions.

  3. The experiments cover several PDE problems, including nonlinear equations, 3D problems, and inverse parameter estimation. The reported speedups over PINNs and CRVPINNs are promising.

Critical weaknesses
  1. The methodological novelty and advantages over existing numerical methods need further clarification. For linear PDEs, the proposed approach reduces to classical least-squares spline collocation, while its extension to nonlinear problems uses established Gauss-Newton optimization. A more detailed discussion and experimental comparison with existing methods, such as isogeometric least-squares collocation (Lin et al., 2020) and isogeometric collocation (Auricchio et al., 2010), would help establish the contribution of the proposed framework.

  2. The theoretical analysis contains a gap. Eq (12) assumes an upper bound on the discrete sampling norm, but Eq (20) uses the reverse inequality. The convergence proof therefore appears to require an additional sampling stability assumption that is not established in the current analysis.

Lin, Hongwei, et al. "Isogeometric least-squares collocation method with consistency and convergence analysis." Journal of Systems Science and Complexity 33.5 (2020): 1656-1693.

Auricchio, Ferdinando, et al. "Isogeometric collocation methods." Mathematical Models and Methods in Applied Sciences 20.11 (2010): 2075-2107.

Questions for authors

In Table 4, the Allen-Cahn solver times for 40 and 80 elements are identical for both spline degrees, whereas Figure 6 shows substantially different runtimes. Could the authors clarify this discrepancy?

Reviewer_p33mOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes IGA-ODIL, which replaces neural networks with B-spline representations and minimizes PDE residuals using least-squares or Gauss–Newton solves. Experiments on forward and inverse PDE problems report speedups over PINN-based methods.

Critical strengths
  • The paper addresses an important direction: fast physics-informed training through analytic derivatives, and least-squares or Gauss–Newton solves.
Critical weaknesses
  • I believe the paper insufficiently developed, with substantial shortcomings in mathematical rigor and clarity. The central idea and contributions are not clearly articulated.

  • The methodological description lacks precise notation, including dimensions of key quantities. The exposition focuses on a two-dimensional special case without clearly presenting the general formulation, leaving the method’s scope unclear despite the three-dimensional experiment.

  • Theorem 1 makes essentially no contribution. It assumes the key properties that should be proved, making the result trivial and uninformative.

Questions for authors

I believe the paper requires substantial revision beyond what a rebuttal can address. Unless the authors demonstrate a factual error in my assessment, a defense of their perspective will not change my recommendation.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.