Learning to Solve PDEs on Neural Shape Representations

AI summary

Overview

• Research area: Machine learning for physics simulation / geometry processing — specifically, learned solvers for partial differential equations (PDEs) defined on 3D surfaces.
• Technical level: Advanced. The paper assumes familiarity with surface PDEs, the Closest Point Method, finite elements, neural implicit shape representations, and neural operators.
• Scope (one sentence): The paper proposes a mesh-free, geometry-conditioned learned operator that solves surface PDEs (heat diffusion and Poisson) directly inside neural shape representations, trained once on a single exemplar shape and generalized across geometries, topologies, and input modalities.

What This Paper Is About

Solving PDEs on surfaces is central to geometry processing and engineering, but classical solvers (finite element methods, the Closest Point Method) require polygonal or triangle meshes. Modern 3D assets increasingly exist as neural shape representations — point clouds, splats, neural implicit fields, overfitted implicit shapes — which are differentiable and topology-agnostic, creating a mismatch where mesh-centric solvers cannot operate in the domain where the data lives. The paper's goal is to learn a local, geometry-conditioned update operator that performs the surface-to-ambient extension step of embedding-based solvers implicitly, so that surface PDEs can be solved directly on neural representations without mesh extraction, per-instance residual training, or extend–restrict shuttling.

Key Contributions

1. A mesh-free, end-to-end differentiable solver for surface PDEs that operates directly on both neural and classical surface representations, avoiding explicit meshing and round-trip errors.
2. A lightweight, shape-conditioned neural operator trained on a single shape (the "Spike" surface) that implicitly learns the local narrow-band extension, requires no per-shape optimization, and can act as a differentiable PDE layer inside existing training pipelines.
3. Extensive evaluation showing generalization across unseen shapes, topologies, input functions, and shape representations, with competitive accuracy and speed on the heat diffusion and Poisson equations, benchmarked against surface FEM, the Closest Point Method (CPM), and GINO.
4. A patch-based local decomposition with a coverage condition relating band width ε, grid spacing Δx, and the number of neighbors k (ε ≤ Δx (3k/4π)^(1/3)), enabling controlled overlap and full band coverage.

Main Findings

• Accuracy on the sphere matches CPM. On the analytic Poisson benchmark (unit sphere, ground truth from spherical harmonics), the method's normalized mean absolute error (NMAE) at the "very fine" resolution is 1.32×10⁻² versus CPM's 1.48×10⁻², while SFEM reaches 1.11×10⁻⁴. The authors state the solver "matches CPM in accuracy" and that comparisons to FEM are grounding rather than competition, since FEM solvers are mature and highly optimized.
• Runtime is competitive with CPM under a shared band solver. Because CPM and the proposed method share the same band-based PDE solvers, runtime differences isolate the extension step. On Poisson on the sphere, "very fine" resolution: SFEM 5.46 s, CPM 335.0 s, Ours 72.6 s. At coarse resolution: SFEM 0.06 s, CPM 10.5 s, Ours 5.10 s.
• Generalization to unseen ShapeNet shapes. Trained only on the Spike, the model ("spike-only") is evaluated on the Poisson equation over five unseen shapes (A–E; chairs and tables), reporting NRMSE scaled by 10⁻². Ours: 1.05, 0.909, 0.986, 0.354, 0.364. CPM: 1.92, 0.892, 0.936, 0.889, 0.600. GINO (spike-only): 13.59, 13.89, 8.58, 6.67, 7.38. GINO with per-shape overfitting achieves the best numbers (0.33, 0.21, 0.34, 0.16, 0.18), but the authors report that they consistently outperform GINO on the same train–test split, indicating genuine generalization to new data.
• Trained on polynomials, generalizes to unseen functions. Supervision uses monomials up to total degree 5. Despite this, the method accurately captures heat diffusion dynamics for a high-frequency initial condition, demonstrating robustness beyond the training distribution.
• Stable under remeshing and connectivity changes. The authors report that errors are notably stable under remeshing and connectivity changes (with details deferred to the supplemental), indicating reduced sensitivity to sampling irregularities and local topology.
• Boundary handling follows CPM practice. Neumann conditions are naturally satisfied because the learned update enforces normal consistency; for exterior Dirichlet boundaries, boundary values are clamped and only the band is updated. On the Max Planck head, heat with homogeneous and sinusoidal Dirichlet data produces errors that "remain low and stable" against a high-resolution SFEM reference.
• Single training run serves multiple PDEs. One operator, trained once on the Spike, is used for both the heat diffusion and Poisson benchmarks across the evaluated representations.

Methodology in Plain English

The starting point is the Closest Point Method (CPM), which embeds a surface PDE into a thin Cartesian "narrow band" around the surface and solves a volumetric PDE there with ordinary finite differences. For the band solution to match the surface solution, the field must be constant along the surface normals — this is called the closest point extension. Classical CPM alternates a solve step with an explicit re-extension step that requires closest-point queries.

This paper replaces the explicit re-extension with a small learned network. The surface is covered with overlapping, surface-centered patches. Each patch records a local frame (the surface point, its normal, and its two principal curvature directions), the nearby band grid nodes, and nearby surface samples with normals. Everything is expressed in the local frame, making the representation invariant to translation and rotation. A compact network composed of three small MLPs plus a learnable scalar consumes the query point, the local band nodes, the surface features, and the current field values, and outputs an updated field value at that query — a single geometry-conditioned grid-to-grid step.

Local predictions are blended into a global band field using proximity-weighted averaging with a temperature parameter T. The blended field is then advanced with standard finite differences and forward Euler (for heat, Δt = 0.1 Δx²), and the aggregate–evolve cycle repeats. The band field is read out on the surface at any time via radial basis function interpolation (Gaussian kernels).

Training uses a single shape — the Spike, represented as a spherical neural surface — since the operator depends only on first- and second-order quantities (normals and curvatures). For each patch, the ground truth is the value of monomial functions at the closest points. Two losses are combined: a mean squared error for function reconstruction, and a normal-consistency term that penalizes the component of the field gradient along surface normals (computed by automatic differentiation). The total objective is L = L_MSE + α·L_NC, with α chosen by ablation.

The approach is representation-agnostic. Geometric cues are derived per modality: mesh normals from area-weighted face averages and curvatures from discrete differential operators; point cloud normals from PCA over k-NN neighborhoods; spherical neural surfaces via differentiation of the mapping (Weingarten map); implicit SDF fields via ∇φ/‖∇φ‖ with curvature directions from the tangent-projected Hessian eigenvectors; and Gaussian splats treated as a filtered point cloud.

Why This Matters

For research, this work closes a gap between mature mesh-based PDE solvers and the neural shape representations that modern 3D pipelines actually use. It removes the need to extract a mesh, solve, and map results back — a round trip that breaks differentiability and adds engineering overhead — while also avoiding the per-instance optimization and poor cross-shape generalization of surface PINNs. Because the learned operator only models the geometric extension step, it stays PDE-agnostic and can drop into existing training pipelines as a neural PDE layer.

Real-world applications suggested by the paper's framing:

• Shape analysis and geometry processing tasks such as heat flow on surfaces, the Poisson equation, and harmonic interpolation.
• Engineering simulation where PDEs must be solved on complex surface geometries.
• Inverse problems and PDE-prior workflows, where end-to-end differentiability through the solver is essential.
• Generative and reconstruction pipelines built on neural implicit fields, point clouds, splats, or neural surfaces, which currently lack a native surface-PDE primitive.

Industry relevance stems from the cost profile: the method needs no meshing and no per-instance optimization, and GINO-style alternatives require retraining for each new PDE setting — the paper reports approximately 8 hours per retraining. A single trained operator reused across PDEs and unseen shapes is a materially cheaper workflow.

Future Directions

• Characterizing failure modes. The provided content is truncated and does not include an explicit limitations section; the range of shapes and topologies where the local operator degrades is not reported.
• Scaling and performance envelope. The method is compared to FEM as "grounding rather than competition," and at the very fine sphere resolution SFEM completed in 5.46 s versus 72.6 s for the proposed method. Whether the gap narrows at other resolutions or for other PDEs is not reported.
• Beyond closed surfaces. Evaluations focus on closed surfaces plus a Dirichlet case on the Max Planck head; behavior on open surfaces, sharp features, and evolving or noisy geometry is left to future work.
• Broader PDE families and modalities. Only heat diffusion and Poisson are benchmarked, across meshes, point clouds, spherical neural surfaces, overfitted occupancy fields, Gaussian splatting, and deep implicit fields. Extending to other operator types and representations is an open question the setup invites.

Target Audience

Researchers and practitioners working at the intersection of machine learning and geometry processing: neural operator and physics-informed learning groups, computational geometry and graphics researchers, and engineers who need surface PDE solutions on modern neural 3D assets. Readers should be comfortable with surface differential operators, finite element discretization, embedding-based solvers, and implicit neural representations; the paper's baseline discussion (SFEM, CPM, GINO) is written for an audience already familiar with that landscape.

Authors’ abstract

Solving partial differential equations (PDEs) on shapes underpins many shape analysis and engineering tasks; yet, prevailing PDE solvers operate on polygonal/triangle meshes while modern 3D assets increasingly live as neural representations. This mismatch leaves no suitable method to solve surface PDEs directly within the neural domain, forcing explicit mesh extraction or per-instance residual training, preventing end-to-end workflows. We present a novel, meshfree formulation that learns a local update operator conditioned on neural (local) shape attributes, enabling surface PDEs to be solved directly where the (neural) data lives. The operator integrates naturally with prevalent neural surface representations, is trained once on a single representative shape, and generalizes across shape and topology variations, enabling accurate, fast inference without explicit meshing or per-instance optimization while preserving differentiability. Across analytic benchmarks (heat diffusion and Poisson equations on the sphere) and on diverse shapes and neural surface representations, our method achieves accuracy comparable to classical solvers while enabling a unified, end-to-end pipeline across neural and traditional surface representations. Our source code and project page: https://welschinger.github.io/Learning-to-Solve-PDEs-on-Neural-Shape-Representations/.