betaFeedback

WorksComputer Science

Learnable Splines for Unstructured Neural Operators

Abstract

Neural operators approximate solution operators of partial differential equations at a fraction of the cost of numerical solvers. Efficient grid-based architectures such as the Fourier neural operator (FNO) assume uniform rectangular grids, whereas engineering simulations use large, irregular, and locally refined meshes. We introduce L-SplineFNO, a memory-efficient neural operator that couples an FNO backbone with a sparse, learnable tensor-product B-spline transfer between unstructured meshes and a latent Cartesian grid. A local normalised scatter avoids the geometry-dependent solve of analytic spline transformation while preserving compact support, partition of unity, and spline smoothness by construction. The neural operator, conditioned on global design and operating parameters, learns the mapping of geometry to the field solution. Across six three-dimensional engineering benchmarks, L-SplineFNO achieves strong predictive accuracy while requiring substantially less memory and compute than baseline operators, scaling to 26.6 million nodes in training and 46 million in inference on a single A100 GPU. It also supports zero-shot super-resolution and efficient closed-form spatial derivatives. Uniquely, the analytic spline decoder yields rigorous prediction bounds over continuous space and uncertain parameters, via Bernstein convex hulls and Taylor models, demonstrated on a 29.2-million-node MAST-U tokamak model.

Introduction

Engineering design and qualification rest on repeated numerical simulation, evaluating a system’s governing equations across a range of parameters and conditions (Winsberg & Alvarado, 2026; Lavin et al., 2021). In steady-state problems such as thermal management, structural loading and aerodynamic pressures, the quantity of interest is a field defined over an unstructured finite-element or finite-volume discretisation of a three-dimensional domain, determined by a small set of design and operating parameters (Perreman et al., 2020; Weller et al., 1998). Design sweeps, sensitivity studies and uncertainty quantification each require this map to be evaluated thousands of times, while a single high-fidelity solve can occupy a compute cluster for hours (Keyes et al., 2013). The cost of the solver therefore limits how much of a design space can be explored (Gopakumar et al., 2025). Safety-critical settings such as fusion devices also require guaranteed prediction ranges for mesh nodes under uncertain operating conditions.

Neural operators provide a route past this bound, learning mappings between function spaces with parameters shared across discretisations and approximating solution operators of partial differential equations (PDEs) at a fraction of the cost of the solver that generated their training data [3] (Kovachki et al. 2023; Azizzadenesheli et al. 2024). The Fourier neural operator (FNO) (Li et al., 2021) has been deployed as a surrogate across fluid dynamics (Wen et al., 2022), heat transfer (Yuan et al., 2025) and plasma modelling [5] (Pamela et al. 2025), reporting speed-ups of several orders of magnitude while retaining discretisation convergence and zero-shot super-resolution (Gopakumar et al., 2024). This efficiency follows from the fast Fourier transform, which assumes the field is sampled on a uniform rectangular grid. However, engineering meshes are often irregular, feature-refined and vary across designs, rarely satisfying this assumption. Existing approaches either process the unstructured mesh directly, which can become memory-intensive at large node counts, or preserve structured processing by transferring the field onto a latent grid. Retaining information from the original mesh while keeping the representation memory-efficient is therefore a key challenge in neural PDEs for large unstructured grids. Existing learned transfers use global coordinate deformations (Li et al., 2023a) or kernel integrals over a radius-limited neighbourhood graph (Li et al., 2023b), but neither preserves structure by construction (§ 2).

Our approach: We build the transfer as a sparse, differentiable and structurally constrained linear map from tensor-product B-splines (Piegl & Tiller, 1996; Prautzsch et al., 2002). Compact support gives each mesh point a local stencil whose size is set by the spline degree, not by a search radius. Partition of unity preserves constants. Spline smoothness gives closed-form evaluation and differentiation at any off-grid query point. We call a transfer with all three properties structure-preserving. The same structure gives a non-negative Bernstein basis on each decoder cell, so the prediction lies within the convex hull of its coefficients; propagating a Taylor model of the parameters through the network extends this to rigorous bounds over continuous space and uncertain parameters. The resulting sparse transfer preserves physical locality and gives a memory-efficient link between unstructured meshes and structured Fourier processing.

To further address the parametric nature of engineering problems, we incorporate global design and operating parameters into the operator. With no initial condition to propagate in steady-state problems, the geometry Ω\Omega and a conditioning vector μ\boldsymbol{\mu} of design variables, loads, material properties, and operating conditions carry all sample-specific information. How μ\boldsymbol{\mu} should enter an operator is unsettled, and its effect is confounded by a redundancy that is rarely controlled for (Koupaï et al., 2024). When μ\boldsymbol{\mu} restates geometric quantities already visible in the point cloud, a model can ignore the conditioning pathway and still appear conditioned. Isolating this pathway therefore requires settings in which μ\boldsymbol{\mu} carries information that the mesh cannot provide, up to the limiting case of a fixed mesh on which μ\boldsymbol{\mu} is the only sample-specific signal. Our benchmarks span this range, from the heat sink, where μ\boldsymbol{\mu} restates the geometry, to HIVE, whose fixed mesh leaves μ\boldsymbol{\mu} as the only sample-specific signal (§ 4). We combine this conditioning with the learnable spline transfer to form a complete conditional L-SplineFNO for parametric fields on complex unstructured meshes. Our contributions are:

  • Learnable spline transfer for unstructured neural operators. We introduce L-SplineFNO, which couples an FNO backbone with a learnable, structure-preserving spline transformation from unstructured meshes to a structured latent grid. This replaces the geometry-dependent linear solve of analytic SplineFNO with a learnable, structure-preserving B-spline scatter onto a latent FNO grid, with analytic reconstruction at arbitrary query points.

  • An end-to-end conditional operator framework. We condition the operator on global design and operating parameters through an affine transformation (Perez et al., 2018). The FiLM-style conditioning produces a per-channel scale and shift, applied to the encoded geometry features before the processor backbone. The conditioning method sits in its own module with interchangeable backbones.

  • Rigorous spatial and parameter bounds. Bernstein convex hulls of the spline decoder and second-order Taylor models of the trained network enclose the prediction over continuous space and uncertain parameter boxes, without reference solutions or retraining.

  • Comprehensive evaluation on complex engineering problems. Across six three-dimensional thermal, structural, and aerodynamic benchmarks, we compare L-SplineFNO against representative mesh-native, latent-attention, and gridded neural operators under a common conditioning pipeline and at matched parameter counts, demonstrating strong predictive accuracy while retaining the efficiency of structured latent processing. We further demonstrate linear scaling with mesh size, inexpensive analytic spatial derivatives, and deployment on a 29.2-million-node mesh of the MAST-U tokamak.

Background and Related Work

Conditional operator learning. Let U\mathcal{U} denote a space of solution fields on a domain Ω\Omega, and let μ∈P⊂Rk\boldsymbol{\mu}\in\mathcal{P}\subset\mathbb{R}^{k} collect design variables, material properties, boundary conditions, or operating conditions. For the steady problems considered here, with no input field, the solution operator maps the geometry and parameters to the solution,

N:(μ,Ω)⟼y∈U,Lμ,Ω(y)=f,(1)\mathcal{N}:(\boldsymbol{\mu},\Omega)\longmapsto y\in\mathcal{U},\qquad\mathcal{L}_{\boldsymbol{\mu},\Omega}(y)=f, \tag*{(1)}

where Lμ,Ω\mathcal{L}_{\boldsymbol{\mu},\Omega} is the governing operator and ff the forcing, if any. All sample-specific information enters through Ω\Omega and μ\boldsymbol{\mu}, so the learning task is both geometric and conditional.

Neural operators. Neural operators approximate maps such as Equation (1) using architectures whose parameters are shared across discretisations, so that a trained model converges across mesh discretisations rather than to a fixed resolution [3] (Kovachki et al. 2023). The FNO parametrises the kernel integral as a global convolution in Fourier space ([12] (Li et al. 2021)): for a lifted feature field vtv_t, a layer computes

vt+1(x)=σ(Wvt(x)+F−1(Rϕ⋅F(vt))(x)),(2)v_{t+1}(x)=\sigma\left(Wv_t(x)+\mathcal{F}^{-1}\left(R_{\phi}\cdot\mathcal{F}(v_t)\right)(x)\right), \tag*{(2)}

where F\mathcal{F} is the Fourier transform, RϕR_{\phi} is a learned multiplier on a truncated band of modes, and WW is a pointwise linear map. Physics-informed neural operators further constrain the learned solution family by penalising PDE residuals ((Li et al., 2023c; Ganeshram et al., 2026)). Evaluating F\mathcal{F} with the fast Fourier transform offers only O(Nlog⁡N)O(N\log N) complexity, making the Equation (2) efficient and FNOs a popular choice ([11]), but restricts its operation to uniform rectangular grids.

Operator learning on unstructured geometries. Extensions to unstructured grids fall into three groups. Mesh-native models attend over mesh points or learned groupings of them, as in GNOT (Hao et al., 2023), Transolver (Wu et al., 2024) and its parallel successor Transolver++ (Luo et al., 2025), or compress the point set into latent tokens, as in UPT ([1]); they avoid gridding, but cost and fidelity depend on how aggressively the mesh is compressed. Latent-grid models keep a structured backbone by transferring the data onto a regular domain, through learned deformations in Geo-FNO (Li et al., 2023a), balanced non-uniform partitions in NUNO (Liu et al., 2023), graph-kernel encoders in GINO (Li et al., 2023b), multi-scale attentional encoders with a vision-transformer processor in GAOT (Wen et al., 2026), or optimal-transport maps for surfaces (Li et al., 2025); here the transfer becomes part of the approximation, since coarse grids lose thin features and interfaces, while dense grids and large neighbourhood graphs cost memory. Precomputed representations such as PGD-NO build geometry tokens ahead of time, saving memory at very large mesh sizes but removes the representation from end-to-end training (Zhong et al., 2026). Industry-scale comparisons find no representation uniformly best across accuracy, cost, memory and deployment (Zhong et al., 2025).

B-splines in scientific machine learning. Tensor-product B-splines are compactly supported, smooth, and form a partition of unity (Piegl & Tiller, 1996). Their sparse Jacobians have been used as differentiable layers for geometry reconstruction and finite-element learning ([7]); physics-informed B-spline networks predict control coefficients directly, so that derivatives and boundary conditions can be imposed analytically (Wang et al., 2026); and Neural Spline Operators extend this coefficient-learning view from finite-dimensional PDE parameters to functional inputs (Wang et al., 2025). In each, B-splines serve as the output representation, rather than as a parameter-efficient, learnable transfer across an unstructured mesh and a latent structured grid.

Conditioning on physical parameters. The parameters μ\mu either restate information already visible in the geometry or encode quantities the geometry cannot reveal, such as loads, materials and operating conditions. Broadcasting μ\mu as an extra input channel is simple, but asks the backbone to wastefully preserve and propagate a signal that is spatially constant across the domain (Takamoto et al., 2023). Alternatives inject μ\mu into the network directly: feature-wise linear modulation (FiLM) applies parameter-dependent, channel-wise affine transformations (Perez et al., 2018); adaptive normalisation and cross-attention modulate intermediate features in a similar spirit (Peebles & Xie, 2023; Rebain et al., 2022); and hypernetworks generate the operator’s weights from μ\mu (Jing et al., 2026).

Spline-Gridded Fourier Neural Operators

Figure 1 summarises the method: sparse spline operators carry mesh features to a uniform latent grid for a standard FNO, and reconstruct the output at arbitrary query points. We first introduce conditional input lifting, then formulate analytically SplineFNO, and finally develop L-SplineFNO by replacing its geometry-dependent spline transform with a local learnable scatter.

Image not extracted

Figure 1. End-to-end conditional L-SplineFNO architecture. The unstructured geometry and global conditioning vector are first mapped to conditioned nodal features through coordinate normalisation, positional encoding, and FiLM modulation. A learnable Cp−1C^{p-1} spline scatter then transforms these features from the unstructured mesh to a structured latent grid, where they are processed by a Fourier neural operator (FNO) backbone. The resulting grid representation is mapped back to the physical mesh using analytic spline reconstruction and a pointwise projection to produce the predicted field. This construction provides a sparse, structure-preserving, differentiable mesh-to-grid-to-mesh mapping with the structured Fourier processing.

Notation. Let {wi}i=1N⊂Ω\{w_i\}_{i=1}^{N}\subset\Omega denote the nodes of an unstructured input mesh in a dd-dimensional domain Ω⊂Rd\Omega\subset\mathbb{R}^{d}. We introduce a uniform tensor-product control grid with multi-index set G=∏ℓ=1d{0,…,Sℓ−1}\mathcal{G}=\prod_{\ell=1}^{d}\{0,\ldots,S_{\ell}-1\}; a multi-index is written g=(g1,…,gd)∈Gg=(g_1,\ldots,g_d)\in\mathcal{G}, where gℓg_{\ell} is the grid index along coordinate axis ℓ\ell. The grid contains G:=∣G∣=∏ℓ=1dSℓG := \lvert G\rvert= \prod_{\ell=1}^{d} S_{\ell} control points, with coordinates ξg(ℓ)=ξ0(ℓ)+gℓΔℓ\xi_{g}^{(\ell)} = \xi_{0}^{(\ell)} + g_{\ell}\Delta_{\ell}, ℓ=1,…,d\ell= 1,\ldots,d, where ξ0(ℓ)\xi_{0}^{(\ell)} and Δℓ>0\Delta_{\ell} > 0 are the origin and spacing along axis ℓ\ell. The spacing may differ between coordinate axes, but is constant along each axis.

Conditional Input Lifting

Many computational physics problems depend on both geometry and global parameters such as material properties, boundary conditions, source terms, or operating conditions. We introduce conditioning in our framework at the input lifting stage to incorporate these problem-specific parameters into the nodal features supplied to the operator. The model receives an unstructured set of geometry coordinates {ωi}i=1N\{\omega_i\}_{i=1}^{N}, a global conditioning vector μ∈Rk\mu\in\mathbb{R}^{k}, and a set of output query points {qj}j=1Nq\{q_j\}_{j=1}^{N_q}. No input field is supplied, and the sample-specific signal is carried by the geometry and μ\mu.

In the feature path, each normalised coordinate ωˉi∈[0,1]d\bar{\omega}_i \in[0,1]^d is encoded by integrated sinusoidal features φ(ωˉi)∈R2dnf\varphi(\bar{\omega}_i) \in\mathbb{R}^{2dn_f} at nfn_f frequencies. The global condition is embedded as eμ=Eμ(μ)e_{\mu} = E_{\mu}(\mu) and applied through FiLM, the most efficient of the conditioning methods we compared (Section G),

f~i=γμ⊙φ(ωˉi)+βμ,hi=Pin[f~i;ωˉi],(3)\tilde{f}_i = \gamma_{\mu} \odot\varphi(\bar{\omega}_i) + \beta_{\mu}, \qquad h_i = P_{\mathrm{in}}\left[\tilde{f}_i;\bar{\omega}_i\right], \tag*{(3)}

where γμ=Hγ(eμ)\gamma_{\mu} = H_{\gamma}(e_{\mu}) and βμ=Hβ(eμ)\beta_{\mu} = H_{\beta}(e_{\mu}) are broadcast across all mesh nodes. The modulation acts only on the encoded features φ\varphi; the normalised coordinates bypass it and are concatenated back before the pointwise lift. We collect the lifted nodal features as X=[h1,…,hN]T∈RN×cX = [h_1,\ldots,h_N]^{\mathsf{T}} \in\mathbb{R}^{N \times c}. In the geometry path, the same normalised coordinates ωˉi\bar{\omega}_i, unmodulated, determine the compact mesh-to-grid spline stencils (Figure 1).

SplineFNO

After constructing the conditioned nodal features, we first introduce SplineFNO, which uses fixed analytic spline operators to transform from the unstructured mesh to a latent uniform grid and to reconstruct at the output coordinates, with a standard FNO operating on the latent grid. Writing ti(ℓ)=(ωi(ℓ)−ξ0(ℓ))/Δℓt_i^{(\ell)} = (\omega_i^{(\ell)} - \xi_0^{(\ell)})/\Delta_{\ell} for the continuous grid coordinate of mesh node ωi\omega_i along axis ℓ\ell, and letting BpB_p denote the centred cardinal B-spline of degree pp introduced in Section B.1, we use p=3p = 3 by default and write B:=BpB := B_p for conciseness.

The unnormalised tensor product basis matrix Φ~∈RN×G\widetilde{\Phi} \in\mathbb{R}^{N \times G} is defined as

Φ~i,g=∏ℓ=1dB(ti(ℓ)−gℓ),i=1,…,N,g∈G.(4)\widetilde{\Phi}_{i,\mathbf{g}}=\prod_{\ell=1}^{d}B\left(t_i^{(\ell)}-g_\ell\right),\qquad i=1,\ldots,N,\quad\mathbf{g}\in\mathcal{G}. \tag*{(4)}

Because the degree pp spline has compact support over p+1p+1 grid points, each row of Φ~\widetilde{\Phi} contains at most (p+1)d(p+1)^d nonzero entries. To preserve constants near the boundary, contributions outside G\mathcal{G} are discarded and the remaining basis values are normalised row wise as Φi,g=Φ~i,g/∑k∈GΦ~i,k\Phi_{i,\mathbf{g}}=\widetilde{\Phi}_{i,\mathbf{g}}/\sum_{\mathbf{k}\in\mathcal{G}}\widetilde{\Phi}_{i,\mathbf{k}}, so that partition of unity is maintained: ∑g∈GΦi,g=1\sum_{\mathbf{g}\in\mathcal{G}}\Phi_{i,\mathbf{g}}=1, whenever the spline support intersects the control grid.

The forward pass consists of three stages. (i) Spline transformation: unstructured to structured. We seek grid coefficients XG∈RG×c\mathcal{X}_{\mathcal{G}}\in\mathbb{R}^{G\times c} whose spline reconstruction approximates the nodal features,

XG=(ΦTΦ+λIG)−1ΦTX,(5)\mathcal{X}_{\mathcal{G}}=\left(\Phi^{\mathsf{T}}\Phi+\lambda I_{\mathcal{G}}\right)^{-1}\Phi^{\mathsf{T}}\mathcal{X}, \tag*{(5)}

where λ>0\lambda>0 ensures that ΦTΦ+λIG\Phi^{\mathsf{T}}\Phi+\lambda I_{\mathcal{G}} is positive definite. (ii) Grid-space neural operator. The coefficients are reshaped according to the tensor-product grid and processed by a standard FNO, YG=Fθ(XG)\mathcal{Y}_{\mathcal{G}}=\mathcal{F}_{\theta}(\mathcal{X}_{\mathcal{G}}). (iii) Spline reconstruction: structured to unstructured. The grid features are evaluated using the same analytic basis,

Y=ΦFθ((ΦTΦ+λIG)−1ΦTX).(6)\mathcal{Y}=\Phi\mathcal{F}_{\theta}\left(\left(\Phi^{\mathsf{T}}\Phi+\lambda I_{\mathcal{G}}\right)^{-1}\Phi^{\mathsf{T}}\mathcal{X}\right). \tag*{(6)}

The spline matrices are sparse, analytic functions of the sampling coordinates and contain no learned parameters. Learning occurs in Fθ\mathcal{F}_{\theta}, while gradients propagate through both spline transfers.

Learnable SplineFNO (L-SplineFNO)

The spline transformation in Equation (5) requires a geometry-dependent linear solve, whose inverse is generally dense. L-SplineFNO replaces this solve with a local, learnable scatter that directly transfers nodal features to the latent grid. The learned quantities are the transfer kernel shapes, while the grid nodes and analytic spline basis remain fixed. Along each coordinate axis ℓ\ell, we use

κθ(ℓ)(t)=B(t)+∑j=1Jζj(ℓ)B(t−τjh),ζ(ℓ)=V0αθ(ℓ).(7)\kappa_{\theta}^{(\ell)}(t)=B(t)+\sum_{j=1}^{J}\zeta_{j}^{(\ell)}B\left(\frac{t-\tau_j}{h}\right),\qquad\zeta^{(\ell)}=V_0\alpha_{\theta}^{(\ell)}. \tag*{(7)}

References

  1. [1]Benedikt Alkin, Andreas Fürst, Simon Schmid, Lukas Gruber, Markus Holzleitner, and Johannes Brandstetter. Universal physics transformers: A framework for efficiently scaling neural operators, 2025. URL https://arxiv.org/abs/2402.12365.
  2. [3]Julius Berner, Miguel Liu-Schiaffini, Jean Kossaifi, Valentin Durruisseaux, Boris Bonev, Kamyar Azizadenesheli, and Anima Anandkumar. Principled approaches for extending neural architectures to function spaces for operator learning. Nature Machine Intelligence, 8(7):1173–1181, 2026. ISSN 2522-5839. doi: 10.1038/s42256-026-01267-z. URL https://doi.org/10.1038/s42256-026-01267-z.
  3. [4]Wolfgang Boehm and Andreas Müller. On de casteljau’s algorithm. Computer Aided Geometric Design, 16 (7):587–605, 1999. ISSN 0167-8396. doi: https://doi.org/10.1016/S0167-8396(99)00023-0. URL https://www.sciencedirect.com/science/article/pii/S0167839699000230.
  4. [5]N. Carey, L. Zanisi, S. Pamela, V. Gopakumar, J. Omotani, J. Buchanan, J. Brandstetter, Fabian Paischer, Gianluca Galletti, and Paul Setinek. Neural operator surrogate models of plasma edge simulations: feasibility and data efficiency. Nuclear Fusion, 65(10):106010, sep 2025. doi: 10.1088/1741-4326/adfdfb. URL https://doi.org/10.1088/1741-4326/adfdfb.
  5. [6]I. T. Chapman, S. C. Cowley, and H. R. Wilson. The spherical tokamak for energy production: theme issue introduction. Philosophical Transactions of the Royal Society A: Mathematical, Physical and Engineering Sciences, 382(2280):20230416, 08 2024. ISSN 1364-503X. doi: 10.1098/rsta.2023.0416. URL https://doi.org/10.1098/rsta.2023.0416.
  6. [7]Minsu Cho, Aditya Balu, Ameya Joshi, Anjana Deya Prasad, Biswait Khara, Soumik Sarkar, Baskar Ganapathy subramanian, Adarsh Krishnamurthy, and Chinmay Hegde. Differentiable spline approximations, 2021. URL https://arxiv.org/abs/2110.01532.
  7. [8]G. Cunningham. High performance plasma vertical position control system for upgraded mast. Fusion Engineering and Design, 88(12):3238–3247, 2013. ISSN 0920-3796. doi: https://doi.org/10.1016/j.fusengdes.2013.10.001. URL https://www.sciencedirect.com/science/article/pii/S0920379613006753.
  8. [9]Pedro M. P. Curvo, Jan-Willem van de Meent, and Maksim Zhdanov. Mspt: Efficient large-scale physical modeling via parallelized multi-scale attention, 2026. URL https://arxiv.org/abs/2512.01738.
  9. [10]Carl de Boor. A Practical Guide to Splines, volume 27 of Applied Mathematical Sciences. Springer-Verlag, New York, NY, 1 edition, 1978. ISBN 978-0-387-95366-3.
  10. [11]Maarten V. de Hoop, Daniel Zhengyu Huang, Elizabeth Qian, and Andrew M. Stuart. The cost-accuracy trade-off in operator learning with neural networks, 2022. URL https://arxiv.org/abs/2203.13181.
  11. [12]Valentin Duruisseaux, Jean Kossaifi, and Anima Anandkumar. Fourier neural operators explained: A practical perspective, 2026. URL https://arxiv.org/abs/2512.01421.

Paper details

vlnc:W_sgsrarqt
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Computer Science

Not recorded for this paper

View statement
Looking for related papers…
Contents