Fourier Neural Operators for Parametric PDEs: Mathematical Foundations, Spectral Approximation Theory, and Empirical Analysis on Kolmogorov Flow and Navier–Stokes

Most existing Partial Differential Equations (PDEs) involve complex boundary geometries; hence, finding solutions to PDEs manually is nearly impossible. Moreover, classical numerical approaches, such as the finite element method (FEM) or the finite difference method (FDM), are often computationally prohibitive, particularly in settings requiring repeated solves across varying parameters. In contrast, data-driven surrogates, primarily used to solve PDEs known as ‘neural operators’, which recently emerged as a promising alternative to solve PDEs, are thousands of times faster than the tradition ways. This paper presents a compendious study of Fourier Neural Operators (FNOs) practiced across four PDE families: the heat equation, the wave equation, Darcy flow, and the Navier–Stokes equation. I derive the complete FNO architecture from first principles (integral operator formulation, lifting, spectral convolution, and projection) and prove a formal approximation bound kG(a)−Gθ (a)k ≤ Ck−s maxM +ε linking the mode budget kmax to the Sobolev regularity s of the target operator. Two experiments, extracted directly from a Colab notebook implementation, are carried out in my study: (i) an FNO on a 2D Kolmogorov-type dissipative dataset (N = 100, grid 128×128, 80 training samples) achieving training relative L2, loss 0.35, but test error 1.1263 due to a combination of mode-truncation floor, data scarcity, and a data-pairing pathology characterized theoretically; and (ii) an FNO on a Navier–Stokes vorticity dataset (N = 5000, grid 64 ×64, 50 time steps, 4000 training samples) with 2D heatmap, 3D surface, and temporal error visualizations. This paper provides a four-source root-cause decomposition of the generalization gap, closed-form analysis of cross-PDE regularity asymmetry, and principled remediation strategies, predominantly demonstrating that Sobolev-class operator approximation theory is a precise predictive tool for FNO practitioners.