[論文レビュー] Fourier Neural Operator for Parametric Partial Differential Equations

Zongyi Li, Nikola Kovachki+5 more|arXiv (Cornell University)|Oct 18, 2020

Model Reduction and Neural Networks参考文献 32被引用数 1,078

ひとこと要約

本論文は Fourier Neural Operator (FNO) を紹介します。これはメッシュに不変な神経演算子で、フーリエ空間でカーネルをパラメータ化することにより関数空間間の写像を学習し、Burgers、Darcy、および Navier–Stokes のような PDE に対して高速で zero-shot 超解像を実現します。

ABSTRACT

The classical development of neural networks has primarily focused on learning mappings between finite-dimensional Euclidean spaces. Recently, this has been generalized to neural operators that learn mappings between function spaces. For partial differential equations (PDEs), neural operators directly learn the mapping from any functional parametric dependence to the solution. Thus, they learn an entire family of PDEs, in contrast to classical methods which solve one instance of the equation. In this work, we formulate a new neural operator by parameterizing the integral kernel directly in Fourier space, allowing for an expressive and efficient architecture. We perform experiments on Burgers' equation, Darcy flow, and Navier-Stokes equation. The Fourier neural operator is the first ML-based method to successfully model turbulent flows with zero-shot super-resolution. It is up to three orders of magnitude faster compared to traditional PDE solvers. Additionally, it achieves superior accuracy compared to previous learning-based solvers under fixed resolution.

研究の動機と目的

• パラメトリック PDE のために、関数空間から関数空間へ写像するニューラル演算子フレームワークを開発する。
• フーリエ空間でカーネルをパラメータ化することにより、メッシュ不変性とゼロショット超解像を達成する。
• 複数の PDE に対して、従来の解法や従来の学習ベース手法より高速化と精度向上を示す。
• 時間依存および時間非依存の PDE への適用性を示し、データ要件と制約について議論する。

提案手法

• Define an iterative neural operator architecture where v_{t+1}(x) = σ(W v_t(x) + (K(a; φ) v_t)(x)).
• Replace the kernel integral operator with a Fourier-space convolution operator to enable efficient computation via FFT.
• Parameterize the Fourier kernel κ_φ through its Fourier transform R_φ and truncate to k_max modes for tractable learning.
• Use four Fourier integral operator layers with ReLU activations and batch normalization to form the FNO.
• Demonstrate discretization-invariance and zero-shot super-resolution by training on low-resolution data and evaluating on higher-resolution grids.

実験結果

リサーチクエスチョン

• RQ1 Can a neural operator learn the solution operator for parametric PDEs in an input-function to output-function setting?
• RQ2 Does parameterizing the kernel in Fourier space yield mesh-invariant, fast, and accurate operators across different PDEs and resolutions?
• RQ3 Can the Fourier neural operator achieve zero-shot super-resolution and outperform existing neural operators and baselines on Burgers’, Darcy, and Navier–Stokes equations?
• RQ4 What are the data requirements and trade-offs for training FNOs for complex PDEs, and how do they perform in Bayesian inverse problems?

主な発見

• The Fourier neural operator learns resolution-invariant solution operators for Burgers’, Darcy, and Navier–Stokes in turbulent regimes.
• FNO achieves zero-shot super-resolution and can be evaluated at higher resolutions than seen during training.
• On fixed resolutions, FNO outperforms benchmarks by roughly 30% for Burgers’, ~60% for Darcy, and ~30% for Navier–Stokes in relative error.
• On a 256×256 grid, FNO inference time is 0.005 s versus 2.2 s for a pseudo-spectral solver, with no accuracy loss in downstream tasks like Bayesian inference.
• FNO-3D (space-time convolution) often provides best performance when data are sufficient; FNO–2D variants still outperform several baselines when data are limited.
• The model remains effective with non-periodic boundaries and can recover high-frequency content through activations between layers.