Abstract:The computational efficiency of many neural operators, widely used for learning solutions of PDEs, relies on the fast Fourier transform (FFT) for performing spectral computations. As the FFT is limited to equispaced (rectangular) grids, this limits the efficiency of such neural operators when applied to problems where the input and output functions need to be processed on general non-equispaced point distributions. Leveraging the observation that a limited set of Fourier (Spectral) modes suffice to provide the required expressivity of a neural operator, we propose a simple method, based on the efficient direct evaluation of the underlying spectral transformation, to extend neural operators to arbitrary domains. An efficient implementation* of such direct spectral evaluations is coupled with existing neural operator models to allow the processing of data on arbitrary non-equispaced distributions of points. With extensive empirical evaluation, we demonstrate that the proposed method allows us to extend neural operators to arbitrary point distributions with significant gains in training speed over baselines while retaining or improving the accuracy of Fourier neural operators (FNOs) and related neural operators.

What problem does this paper attempt to address?

The main problem that this paper attempts to solve is the limitations of existing Fourier - transform - based neural operators (such as FNO) when dealing with non - uniformly distributed data points. Specifically: 1. **Limitations of existing methods**: - **Limitations of Fourier transform (FFT)**: Existing FNO and other Fourier - transform - based neural operators rely on the fast Fourier transform (FFT), and FFT can only be applied to data points on an equally - spaced (rectangular) grid. This makes these neural operators inefficient when dealing with problems of general non - equally - spaced point distributions. - **Insufficiency of interpolation methods**: In order to apply FNO on non - equally - spaced points, some methods convert the data to an equally - spaced grid through interpolation, but this will lead to an increase in computational cost or a decrease in accuracy. 2. **The method proposed in the paper**: - **Direct spectral transformation**: The paper proposes a method based on direct spectral evaluations (DSE), which can efficiently perform spectral transformation on any point distribution. This method realizes Fourier transformation or other spectral transformations by constructing a matrix, thus avoiding the need for interpolation. - **Computational efficiency**: The DSE method has an advantage in computational complexity. Especially when dealing with a small number of modes, its computational complexity is \(O(mN)\), where \(m\) is the number of modes and \(N\) is the number of data points. In contrast, the traditional FFT method is less efficient when dealing with non - equally - spaced points. 3. **Experimental verification**: - **Multiple tasks**: The paper verifies the effectiveness of the DSE method through multiple experiments, including the one - dimensional Burgers equation, the two - dimensional shear layer problem, surface humidity prediction, flow around airfoils and elasticity problems. - **Performance comparison**: The experimental results show that the DSE method is not only significantly superior to the baseline method in training speed, but also performs well in terms of test error, especially when dealing with data with non - equally - spaced point distributions. In conclusion, this paper aims to overcome the limitations of existing Fourier - transform - based neural operators when dealing with data with non - equally - spaced point distributions by proposing a new direct spectral transformation method, thereby improving computational efficiency and model accuracy.

Beyond Regular Grids: Fourier-Based Neural Operators on Arbitrary Domains

Toward a Better Understanding of Fourier Neural Operators from a Spectral Perspective

Model-Parallel Fourier Neural Operators as Learned Surrogates for Large-Scale Parametric PDEs

Multi-Grid Tensorized Fourier Neural Operator for High-Resolution PDEs

Spectral Neural Operators

Enhancing Solutions for Complex PDEs: Introducing Complementary Convolution and Equivariant Attention in Fourier Neural Operators

Non-equispaced Fourier Neural Solvers for PDEs

Domain Agnostic Fourier Neural Operators

Incremental Spatial and Spectral Learning of Neural Operators for Solving Large-Scale PDEs

Spherical Fourier Neural Operators: Learning Stable Dynamics on the Sphere

Fourier Neural Operator with Learned Deformations for PDEs on General Geometries

MgFNO: Multi-grid Architecture Fourier Neural Operator for Parametric Partial Differential Equations

Spatio-spectral graph neural operator for solving computational mechanics problems on irregular domain and unstructured grid

Discretization Error of Fourier Neural Operators

Resolution-Invariant Image Classification based on Fourier Neural Operators

Bounding the Rademacher complexity of Fourier neural operators

Fourier Neural Operator Surrogate Model to Predict 3D Seismic Waves Propagation

Gabor-Filtered Fourier Neural Operator for Solving Partial Differential Equations

Accelerating the convergence of Newton's method for nonlinear elliptic PDEs using Fourier neural operators

Spectral-Refiner: Fine-Tuning of Accurate Spatiotemporal Neural Operator for Turbulent Flows