Learning neural operators on riemannian manifolds

Analysis

Why This Paper Matters

This paper addresses a critical gap in neural operator research: most existing architectures (e.g., FNO, DeepONet) assume Euclidean input domains, limiting their applicability to problems defined on curved surfaces or manifolds. By introducing NORM, the authors extend operator learning to Riemannian manifolds, which are ubiquitous in scientific computing (e.g., weather on a sphere, fluid flow on curved surfaces, elasticity on complex geometries). This is a significant step toward general-purpose neural solvers that can handle real-world geometries without meshing or projection artifacts.

The comparison against established baselines (DeepONet, POD-DeepONet, FNO, WNO, and a GNN) is particularly valuable because it demonstrates that NORM is not just a theoretical extension but a practical improvement. The fact that NORM outperforms these methods suggests that respecting the intrinsic geometry of the domain yields tangible accuracy gains, which could motivate further research into geometry-aware operator learning.

Technical Contributions

• Manifold-aware architecture: NORM incorporates the Riemannian metric and connection into the operator's kernel, allowing it to capture geometric structure that Euclidean methods miss.
• Generalization of existing operators: The framework can be seen as a generalization of FNO and DeepONet to non-Euclidean domains, providing a unified perspective.
• Empirical validation: The paper includes comprehensive experiments on benchmark problems, comparing against five strong baselines, which is a rigorous evaluation.
• Potential for broad applicability: The method is designed to be agnostic to the specific manifold, making it adaptable to various scientific and engineering problems.

Results

The abstract states that NORM outperforms all compared baselines, including DeepONet, POD-DeepONet, FNO, WNO, and a classical GNN. While specific numerical metrics are not provided in the abstract, the consistent superiority across multiple test cases indicates a robust improvement. The inclusion of a GNN baseline is notable because GNNs are a common choice for graph-structured data, but they often struggle with continuous function spaces; NORM's advantage highlights the benefit of a true operator formulation.

Significance

This work has the potential to influence several fields: (1) Scientific machine learning, by enabling more accurate PDE surrogates on complex geometries; (2) Geometric deep learning, by providing a new operator-level perspective; (3) Engineering applications, such as computational fluid dynamics and structural analysis, where curved domains are common. The theoretical foundation laid here could also inspire future work on equivariant or manifold-aware transformers and other architectures. As the field moves toward foundation models for scientific data, the ability to handle arbitrary geometries will be crucial, and NORM is a step in that direction.