acceptodds
Under review as a conference paper at ICLR 2027

Learning Spectrally Optimised Mesh-Free Discretisations

Abstract

Numerical methods for partial differential equations (PDEs) discretise differential operators, ideally reproducing the action of the continuous operators across all wavenumbers permitted by a given discretisation. Spectral-type methods approach this ideal, but rely on structured grids or high-order meshes that are hard to generate for complex geometries. In contrast, mesh-free methods are geometrically flexible, yet how faithfully they reproduce the operator across resolved scales is strongly influenced by a heuristically chosen kernel, which selects one of many weight sets satisfying the same consistency conditions without regard to the resulting spectral response. To address this, we introduce Spectrally optimised Neural Discretisations (SpeND), a framework that learns the map from local stencil geometry to discretisation weights on unstructured point clouds. A projection layer constrains every predicted stencil to the affine set defined by the discrete moment conditions, so polynomial consistency, and hence formal order of accuracy, holds exactly. Since consistency is enforced by the architecture, the weights can additionally be optimised for accuracy on a prescribed function space. Herein, we use Fourier modes and target the exact differentiation response over a chosen wavenumber band, so training is unsupervised. The trained discrete differential operators are PDE-agnostic and are applied without retraining to the Poisson, Burgers and Navier–Stokes equations, to near-boundary stencils, and across resolutions. At the same order and resolution, SpeND matches or improves on established mesh-free discretisations, and has been shown to reduce the wall-clock time between to achieve equivalent error as the baselines.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
learned discretisation, hybrid PDE solver, mesh-free methods, numerical methods

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_1osiOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a framework for learning mesh-free differential operators on unstructured point clouds. The key idea is to use a neural network to predict discretization weights from local stencil geometries while enforcing polynomial consistency through a moment-projection layer. This projection ensures the prescribed formal order of accuracy, while a Fourier-based training objective optimizes dispersion and dissipation properties over selected wavenumber ranges without requiring PDE solution data.

The authors evaluate SpeND on differential approximation tests, viscous Burgers' equation, Poisson's equation with an embedded obstacle, and weakly compressible Navier–Stokes equations. The results demonstrate improved accuracy and computational efficiency compared with classical RBF-FD and LABFM discretizations, with reported wall-clock speedups of approximately 3–20 times at matched accuracy in the tested regimes.

Critical strengths
  1. combination of exact polynomial consistency and learned spectral optimization.
  2. demonstrate numerical accuracy and efficiency improvements against established mesh-free methods.
Critical weaknesses
  1. The necessity and advantage of neural learning over direct constrained spectral optimization are not sufficiently established.
  2. Stability and robustness on general irregular point clouds remain insufficiently established.
Questions for authors
  1. compare SpeND against a directly spectrally optimized mesh-free discretization, obtained by minimizing the same Fourier-response objective subject to identical moment constraints.
  2. Do the reported advantages persist under strongly anisotropic node distributions or more complex boundaries? Can the authors demonstrate whether the learned operators remain stable under long time integration without additional stabilization beyond that used in the classical baselines?
2 comments

We thank the reviewer for recognising the merits of our paper and for the helpful suggestions. We address each point below; changes are in the revised manuscript.

[W1/Q1]

We thank the reviewer for this suggestion and have added this comparison (in App.G, pages 31-34, highlighted in blue; see also our response to Reviewer zV6j [Q2]). The direct baseline minimises the same Fourier-response objective under identical moment constraints, independently for each stencil. It establishes two advantages of learning:

  • Cost. In SpeND, the optimisation is performed once, offline, during training. The trained PointNet (one per order and operator, independent of the PDE) then applies without retraining to new node distributions of the type it was trained on, whereas direct optimisation must be repeated for every stencil of every new node distribution. We therefore compare SpeND’s inference time against the full direct optimisation; training cost is reported in Appendix J. Matching SpeND’s loss takes direct optimisation 10–18× longer in float64, 17–20× longer in float32 on CPU, and up to 144× longer in float32 on GPU. Running the optimiser to convergence: it attains a median of 67–100× lower per-stencil loss, at 43–48× the construction time on CPU and 25–28× on GPU, in float64.

  • Global behaviour. Stencils optimised in isolation assemble into operators with markedly worse spectra (largest real part for derivative approximately 7 vs 0.9; largest imaginary part for Laplacian 3 vs 0.10). In practice, Burgers diverges at every resolution with the direct weights, also with \Delta t reduced, and BiCGSTAB fails to converge for Poisson at the finest resolution. We believe this demonstrates the advantages of SpeND against direct optimisation.

[W2/Q2] Anisotropic nodes. We have added a Poisson case with a square obstacle (App. F.2.3, pp. 26-29 highlighted in blue), with isotropic nodes and with boundary refinement where the wall spacing is 2× and 4× smaller than the far-field spacing. This tests two configurations absent from training, which contains only stencils truncated by a single straight wall on isotropic nodes:

  • stencils at the corners, truncated by two walls at a right angle;

  • anisotropic stencils near the refined wall, where node spacing varies across the stencil.

All methods retain second-order convergence. With refinement, each method’s error is similar to its isotropic-node error at the same far-field spacing, and SpeND’s speedups over the baselines remain similar to those for the circular obstacle.

Regarding node disorder. The operators are evaluated on particle-shifted clouds and on lattices perturbed with \epsilon = 0.8 (Burgers), well beyond the disorder typical of mesh-free simulations [1]. Ablations extend this to \epsilon = 1, where neighbouring nodes can coincide, and SpeND retains its convergence rate (App. I, Fig. 24, p. 37). The distributions which we have tested, such as disturbed cartesian nodes and particle shifting are standard bench marks for mesh-free discretisation frameworks [1], [2].

For more complex geometries, the limiting factor is the reference solution rather than the method. Our evaluation relies on analytical or manufactured references, so that errors and convergence rates can be measured precisely. In complex geometries or flows no pointwise reference exists, different schemes can agree in their statistics while differing in the instantaneous flow, so neither the error nor the convergence rate can be quantified. The square obstacle extends the tested boundaries as far as a manufactured solution allows, with sharp corners and refinement. Moreover, the main difficulty for a mesh-free discretisation is the irregularity of the node set rather than the shape of the domain, and all our experiments already use fully unstructured nodes. Applications to more realistic geometries and flows are a natural next step, and beyond the scope of the work, and which we now note as future work in the Conclusion (p. 10, highlighted in blue). See also our response to Reviewer FcLX [W1].

Long rollout stability. The learned operators do not use any stabilisation beyond that used in the classical baselines. SpeND is learned discretisation, implemented in a classical framework. The long-time global stability is determined entirely by the eigenspectrum: see the global eigenspectrum analysis in Appendix D, which compares the methods on the same node sets. The stability of SpeND is comparable to RBF-FDs and LABFM. In the Burgers test case (Sec. 4.2, when p = 2), no method uses stabilisation, and the physical viscosity is sufficient for SpeND and LABFM. In the Navier–Stokes cases (p = 4) (Sec 4.4, App F.3, App F.4), all methods use the same hyperviscous filter, with the same damping, applied in the same way, and the same time integrator. With the addition of dealiasing via a hyperviscous filter (note, dealiasing is essential for any high-order collocated method solving the Navier-Stokes equations), all exhibit stable long-time behaviour. No additional damping is added. To test long-time behaviour in a demanding setting, we have added a three-dimensional Taylor–Green vortex at Re = 1600, run to t = 20 through transition to turbulence (Appendix F.4). SpeND remains stable throughout, with the same stabilisation as LABFM. We have included this discussion in Appendix D (highlighted in red).

References:

[1] Suchde, P., Jacquemin, T. and Davydov, O., 2023. Point Cloud Generation for Meshfree Methods: An Overview. Archives of Computational Methods in Engineering, 30(2), pp.889–915. Available at: https://doi.org/10.1007/s11831-022-09820-w.

[2] Quinlan, N. J., Basa, M. and Lastiwka, M. Truncation error in mesh-free particle methods. International Journal for Numerical Methods in Engineering, 66(13):2064–2085, 2006. doi: 10.1002/nme.1617

Reviewer_zV6jOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes SpeND, a framework that learns mesh-free differentiation weights from local point-cloud geometry while enforcing polynomial consistency through an affine projection. Training optimises analytical Fourier responses without requiring PDE solution data, and the learned operators transfer across resolutions and several PDEs. Results show improved accuracy over matched RBF-FD and LABFM configurations and reported speedups. However, the Taylor–Green evaluation uses an incompressible reference for a weakly compressible system and emphasises late-time error floors, while the Burgers comparison lacks sufficient controls to explain RBF-FD’s instability

Critical strengths
  1. Consistency is enforced independently of training. The affine projection satisfies the polynomial moment conditions while leaving the network free to improve the modal response
  2. The Poisson experiment provides convincing evidence of an accuracy–cost benefit. With matched polynomial degree, stencil size, and solver infrastructure, all methods retain second-order convergence around a circular obstacle, while SpeND achieves approximately 7×7\times lower error than RBF-FD at the finest resolution and approximately 8×8\times lower runtime at matched error
Critical weaknesses
  1. The Taylor–Green evaluation does not cleanly establish improved flow accuracy. It uses an incompressible reference for a weakly compressible system without quantifying the resulting discrepancy. It also emphasises late-time error floors based on velocity magnitudes, normalised by an early reference norm
  2. The Burgers comparison lacks a stable RBF-FD baseline. RBF-FD diverges under refinement, but the paper does not specify the time-step rule or establish whether the failures persist with smaller time steps or suitable stabilisation.
Questions for authors
  1. Does SpeND retain its Taylor–Green accuracy–cost advantage when evaluated using vector-velocity errors at common times before substantial decay, against a reference consistent with the weakly compressible equations?
  2. How does SpeND compare with directly minimising the same modal-response objective under the same moment constraints, in both accuracy and stencil-construction cost?
1 comment

We thank the reviewer for recognising the merits of our paper and for the helpful suggestions. We address each point below; changes are in the revised manuscript.

[W1] / [Q1] We redid the evaluation. We now measure errors against a weakly compressible reference eighth-order LABFM, at a resolution of s = 1/256. We observed order of approximately 7, estimated error < 1e-10. The error is computed relative to the L2 norm of the velocity vector, normalised by the reference at the same time, on shared node sets. As suggested by the reviewer, we no longer report late-time error floors, we report it at t ≈ 1 (velocity amplitude about 45% of initial).

At s = 1/128, the errors are 2.98e-7 (SpeND), 6.28e-7 (RBF-FD) and 9.32e-7 (LABFM), all at p = 4. The accuracy ranking of the submission is unchanged, with SpeND the most accurate at s = 1/64 and 1/128. The speed-ups at equal error are revised from 3× to 1.8× over RBF-FD and from 5.5× to 3× over LABFM (App. F.3), to reach equivalent error. Furthermore, in Appendix F.3 (p. 29 - 30) we provide a band decomposition of the errors at t=1. These show that at t=1, SpeND reduces the error at the higher modes by 36–60× against RBF-FD and LABFM, respectively.

[W2] All methods use the same solver, node sets and time-step rule, \Delta t = min(0.1 h/U, 0.05 h^2 Re), the stricter of an advective and a viscous limit, where h is the stencil radius and U = 1 the velocity scale. This is now stated in App. F.1.1 (p. 26, line 1365, highlighted in blue). Repeating the RBF-FD runs with \Delta t reduced by a factor of ½ and ¼, RBF-FD still diverges.

The Burgers case (p = 2) deliberately compares all methods without stabilisation. RBF-FD is known to require hyperviscosity for convective problems [1], while SpeND and LABFM complete these runs without it, which we see as an advantage in robustness. Adding stabilisation to RBF-FD alone would make the comparison unequal. A stabilised comparison is given by the Taylor–Green vortex (Sec. 4.4, with p = 4), where all three methods use the same hyperviscous filter and RBF-FD is stable.

[Q2] We added this baseline in App.G (pages 31-34) highlighted in blue: the same objective and moment constraints, minimised independently for each stencil.

  • Cost at matched loss. Matching SpeND’s loss on each stencil takes direct optimisation 10–18× longer in float64, 17–20× longer in float32 on CPU, and up to 144× longer in float32 on GPU (Table 5). These timings compare SpeND’s inference against the direct optimisation. SpeND’s training is a one-off cost, independent of the PDE, and is reported in Appendix J.

  • Converged direct optimum. Run to convergence, direct optimisation attains a median of 67–100× lower per-stencil loss than SpeND on 1000 test stencils (Table 4), at 43–48× the construction time on CPU and 25–28× on GPU, in float64 (Table 5).

  • Global operators. The lower local loss does not carry over. Optimised in isolation, the stencils assemble into operators with markedly worse spectra: the largest real part for the derivative is approximately 7 vs 0.9 for SpeND, and the largest imaginary part for the Laplacian is 3 vs 0.10 (Fig. 18).

  • PDEs. With the direct weights, Burgers diverges at every resolution, also with \Delta t halved and quartered. In Poisson the direct operator is more accurate at coarse resolutions, but BiCGSTAB fails to converge at the finest, and SpeND reaches every error level faster in wall-clock time.

We believe this demonstrates the advantages of SpeND against direct optimisation. Stability requirements could in principle be added to the direct objective, but this would make it a different problem from the one SpeND solves, which uses no such term, and would likely trade off against the low local loss that is the direct method’s main advantage. Exploring this is beyond the scope of this work. SpeND reaches well-behaved global operators without any explicit stability term, and is trained once and reused on new node distributions of the type it was trained on, whereas direct optimisation must be repeated for every stencil.

References:

[1] Fornberg, B. and Lehto, E., 2011. Stabilization of RBF-generated finite difference methods for convective PDEs. Journal of Computational Physics, 230(6), pp.2270-2285. Available at: https://doi.org/10.1016/j.jcp.2010.12.014.

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

The paper proposes to learn discretization of differential operators on unstructured point clouds using PointNet + additional constraints. The results on several examples show that it gives better convergence that standard methods.

Critical strengths

The idea of learning the stencil of the operators from a PointNet is interesting, as well as proposed scheme that also leaves the approximation order and puts it into the learning process.

Critical weaknesses

The examples are quite simple, two-dimensional and do not require unstructured point clouds (i.e.) look artificial.

Questions for authors
  1. What about other properties of discretization such as stability?
  2. Can we get OOD examples in learning of the coefficients?
2 comments

We thank the reviewer for recognising the merits of our paper and for the helpful suggestions. We address each point below.

[W1] Simplicity of the test cases. The test cases were chosen so that errors are measured against analytical or manufactured references, which allows convergence rates and accuracy to be quantified precisely. For time-dependent nonlinear problems (Burgers, Navier–Stokes), such references are essentially limited to simple periodic domains. In complex geometries or flows, no pointwise reference exists: different schemes can agree in their statistics while differing in the instantaneous flow, so pointwise errors and convergence rates cannot be measured.

Moreover, the main difficulty for a mesh-free discretisation is the irregularity of the node set, not the shape of the domain. All experiments use fully unstructured nodes: particle-shifted distributions, and, for Burgers, lattices perturbed with \epsilon = 0.8, which are considerably more disordered than node sets typical of mesh-free simulations. The comparison therefore directly tests each method on unstructured stencils.

To move beyond two-dimensional flows, we have also added the three-dimensional Taylor-Green vortex (App. F.4, page 30, highlighted in blue). SpeND discretises unstructured planes, combined with finite differences in the third direction, so this is an application to a more complex multiscale flow rather than a claim of fully three-dimensional point clouds. In this setting, SpeND at p = 4 matches the accuracy of LABFM at p = 8.

Introducing a new discretisation and validating it in 2D before applying it to complex geometries and 3D flows follows standard practice for mesh-free schemes [1–5]. We agree that applications to more realistic flows are valuable and have added this as future work in the Conclusion, p. 10 “A natural next step is to …” and have stated this in Limitations, p.35

[Q1] Stability. We analyse stability at two levels.

Local (modal) stability. App. C.3 sets out the stability requirements for derivative and diffusion operators. SpeND’s training loss (Sec. 3.3, Eqs. 7–8) is designed from these requirements: it (i) strongly penalises modal amplification and (ii) penalises numerical dispersion, i.e. amplification in modal phase speed.

Global stability. App. D (Fig. 9) reports the eigenvalues of the assembled derivative and Laplacian operators for all consistency orders. While preparing this response, we identified an inconsistency in the submitted Fig. 9: SpeND’s spectra were computed on particle-shifted node sets, whereas the baselines used noisy node sets. We have recomputed all spectra on identical node sets, separately for shifted and noisy distributions. The same inconsistency affected the modal response in Fig. 8, which we have also updated; the conclusions drawn from both figures are unchanged. All other experiments, including every PDE test case, results in the main text, and timings are unaffected.

Practical stability. In Burgers (Sec. 4.2, where p = 2), all methods use the same node sets and time-step rule with no stabilisation; SpeND and LABFM remain stable at all resolutions, while RBF-FD becomes unstable at most. In the Navier–Stokes cases (p = 4), all methods use the same hyperviscous filter.

We agree that the main text did not point to this analysis clearly. We have added a pointer at the end of Sec. 4 (p. 7), highlighted in blue.

[Q2] Out-of-distribution evaluation. The learned map from stencil geometry to weights is evaluated outside its training distribution in several ways:

  • unseen stencil sizes (App. I, Fig. 25, p. 38);
  • an unseen consistency order (App. I, Fig. 23, p. 37);
  • curved boundaries, whereas training used only straight walls (App. F.2.2; Sec. 4.3);
  • new in the rebuttal: a square obstacle for the Poisson equation (App. F.2.3, Figs. 12–14, p. 26 - 29), whose right-angled truncated stencils do not appear in the training data, and anisotropic stencils present in node refinement.

In function space, training uses only Fourier-mode responses at specific isolated wavenumbers, yet the operators are applied without retraining to all PDEs in Sec. 4.

We agree these results deserve more prominence and now summarise them in the Experiments (Sec. 4, p. 6, highlighted in red), with a pointer to the ablations.

References

[1] Chandhini, G. and Sanyasiraju, Y.V.S.S., 2007. Local RBF-FD solutions for steady convection–diffusion problems. International Journal for Numerical Methods in Engineering, 72(3), pp.352-378. Available at: https://doi.org/10.1002/nme.2024.

[2] Lind, S.J. and Stansby, P.K., 2016. High-order Eulerian incompressible smoothed particle hydrodynamics with transition to Lagrangian free-surface motion. Journal of Computational Physics, 326, pp.290-311. Available at: https://doi.org/10.1016/j.jcp.2016.08.047.

[3] Trask, N. Perego, M and Bochev, P.. A high-order staggered meshless method for elliptic problems. SIAM Journal on Scientific Computing, 39(2):A479–A502, 2017. doi: 10.1137/16M1055992. URL https://doi.org/10.1137/16M1055992.

[4] King, J.R.C., Lind, S.J. and Nasar, A.M.A., 2020. High order difference schemes using the local anisotropic basis function method. Journal of Computational Physics, 415, p.109549. Available at: https://doi.org/10.1016/j.jcp.2020.109549.

[5] Lin, M., Fourtakas, G. and Rogers, B.D., 2025. A novel high-order spectral incompressible smoothed particle hydrodynamics (ISPH) scheme with an immersed boundary method (IBM). Journal of Computational Physics, 540, p.114264. Available at: https://doi.org/10.1016/j.jcp.2025.114264.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of revisions

All changes are highlighted in the revised manuscript: blue by default, with red marking a second change in the same section that came from a different comment. The main concerns and how we addressed them:

  1. Learning vs. direct spectral optimisation (zV6j, 1osi). New App. G adds the requested baseline: the same Fourier objective and moment constraints, minimised per stencil.

    • Matching SpeND's loss takes direct optimisation 10–144× longer, depending on precision and hardware.
    • Run to convergence, it reaches 67–100× lower per-stencil loss, but the assembled operators have worse global spectra: the Burgers simulation diverges at every resolution, and BiCGSTAB fails to converge for Poisson's equation at the finest resolution.
    • SpeND is trained once and reused; direct optimisation must be repeated for every stencil.
  2. Taylor–Green evaluation (zV6j). Re-evaluated against a weakly compressible 8th-order LABFM reference at high resolution (estimated error <1e-10). Errors are now vector-velocity L2 errors at t ≈ 1, normalised at the same time; late-time error floors are removed. The accuracy ranking of the submission is unchanged, with SpeND the most accurate at s = 1/64 and 1/128. The speed-ups at equal error are revised from 3× to 1.8× over RBF-FD and from 5.5× to 3× over LABFM (Sec 4.4; App. F.3).

  3. Fairness of the Burgers baseline (zV6j). The time-step rule is now stated (App. F.1.1), and RBF-FD still diverges with \Delta t halved and quartered. All methods are compared without stabilisation; adding it to RBF-FD alone would make the comparison unequal. RBF-FD is known to need hyperviscosity for convective problems [1], while SpeND and LABFM do not need it in this test. In Taylor–Green, all methods share the same hyperviscous filter and all methods are stable.

  4. Stability (FcLX, 1osi).

    • The stability analysis (App. C.2, D) is now referenced from Sec. 4. App. D adds a discussion of how SpeND’s stability relates to that of the baselines and other meshfree formulations.
    • New 3D Taylor–Green at Re = 1600 to t = 20, through transition to turbulence (App. F.4): SpeND stays stable with the same stabilisation as LABFM, and SpeND at fourth order (p = 4) matches LABFM at eighth order (p = 8).
  5. Generality and OOD behaviour (FcLX, 1osi).

    • New Poisson square-obstacle case (App. F.2.3): corner stencils and anisotropic stencils from 2×/4× wall refinement, neither present in training. Second-order convergence and speed-ups are retained.
    • OOD results (unseen stencil sizes and orders, curved boundaries, corner and anisotropic stencils) are now summarised in Sec. 4, with pointers to the ablations and App. F.2.3.
    • The 3D Taylor–Green vortex (App. F.4) applies SpeND to a more complex multiscale flow. SpeND discretises unstructured planes, with finite differences in the third direction, so this is not a claim of fully 3D point clouds (which we leave for future work).
    • Realistic flows in complex geometries are added as future work in the Conclusion and Limitations. Our test cases were chosen for their analytical or manufactured references, which allow errors and convergence rates to be measured precisely; for chaotic or turbulent flows in complex geometries, no such pointwise reference exists.

Correction. We found that SpeND’s spectra in Fig. 9 and the modal response in Fig. 8 were computed on different node sets from the baselines. Both were recomputed on identical node sets. The recomputed spectra and modal responses are close to the submitted ones, so the conclusions are unchanged, and no other results are affected.

We hope the revisions address the concerns raised and we would like to again thank all reviewers and the AC for their consideration.

References:

[1] Fornberg, B. and Lehto, E., 2011. Stabilization of RBF-generated finite difference methods for convective PDEs. Journal of Computational Physics, 230(6), pp.2270-2285. Available at: https://doi.org/10.1016/j.jcp.2010.12.014.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.