Fourier Neural Operators Explained: A Practical Perspective

TL;DR

This work addresses the challenge of efficiently learning mappings between function spaces for PDEs by providing a practical, theory-grounded guide to Fourier Neural Operators (FNOs) and their implementation in NeuralOperator. It unifies operator theory, spectral methods, and architectural design, detailing how to parameterize, extend, and safely apply FNOs across discretizations, domains, and resolutions. Core contributions include practical guidance on Nyquist-aware mode selection, non-periodic extensions via padding and Fourier continuation, multi-resolution and incremental training (iFNO), physics-informed variants (PINO, PINO), uncertainty quantification (UQNO), and advanced architectures (TFNO, SFNO, Geo-FNO, GINO, OTNO). The guide emphasizes robust hyperparameter tuning, data preprocessing, and visualization practices to enable reliable deployment of FNOs across scientific and engineering domains, including inverse design and multi-physics problems.

Abstract

Partial differential equations (PDEs) govern a wide variety of dynamical processes in science and engineering, yet obtaining their numerical solutions often requires high-resolution discretizations and repeated evaluations of complex operators, leading to substantial computational costs. Neural operators have recently emerged as a powerful framework for learning mappings between function spaces directly from data, enabling efficient surrogate models for PDE systems. Among these architectures, the Fourier Neural Operator (FNO) has become the most influential and widely adopted due to its elegant spectral formulation, which captures global correlations through learnable transformations in Fourier space while remaining invariant to discretization and resolution. Despite their success, the practical use of FNOs is often hindered by an incomplete understanding among practitioners of their theoretical foundations, practical constraints, and implementation details, which can lead to their incorrect or unreliable application. This work presents a comprehensive and practice-oriented guide to FNOs, unifying their mathematical principles with implementation strategies. We provide an intuitive exposition to the concepts of operator theory and signal-processing that underlie the FNO, detail its spectral parameterization and the computational design of all its components, and address common misunderstandings encountered in the literature. The exposition is closely integrated with the NeuralOperator 2.0.0 library, offering modular state-of-the-art implementations that faithfully reflect the theory. By connecting rigorous foundations with practical insight, this guide aims to establish a clear and reliable framework for applying FNOs effectively across diverse scientific and engineering fields.

Figures (27)

• Figure 1: Example of Fourier series for the periodic function $f(x) = \cos(x) + \sin(x)^4 + 0.5\exp(\sin(2x))$ truncated at different values of $N$, next to the $L^2$ error of the truncated Fourier series.
• Figure 2: Power spectrum of the function $f(x) = 3\cos(4x) + 0.1\sin(10x) + \cos(22x)$. Distinct peaks appear at $k = 4$, $10$, and $22$, revealing the dominant periodic components at these frequencies. The amplitude of the peak at $k = 4$ is noticeably higher, reflecting the larger coefficient 3 compared to the weaker contribution from the $k = 10$ term with coefficient 0.1.
• Figure 3: Illustration of aliasing in the time and frequency domains. The continuous signal $f(x) = \cos(20\pi x)$ (in blue) is shown together with its sampled versions at two different sampling rates (in red). When the sampling rate exceeds the Nyquist limit (top), the samples accurately capture the waveform, and the power spectrum shows a single peak at the true frequency. When the sampling rate is below the Nyquist limit (bottom), the samples misrepresent the signal, producing an apparent lower-frequency oscillation in the time domain and a corresponding aliased peak in the frequency spectrum.
• Figure 4: Spectral interpolation. The periodic signal $f(x) = \cos(2x) + \exp(\sin(4x))$ is sampled at $40$ points and its spectrum is computed (top). The signal is then reconstructed at a finer resolution with $120$ points by inserting zeros at the higher frequencies in the spectrum before performing the inverse DFT (bottom).
• Figure 5: Example of periodic extensions of the function $f(x) = \sin(16x) - \cos(8x) + 0.4$ obtained using zero-padding and mirror-padding.

Theorems & Definitions (1)

• Theorem 3.1: Universal Approximation Theorem for FNOs kovachki2021universal