Fourier Neural Operators for Rayleigh-Bénard Convection

2026-07-02 · Source: Machine Learning · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Engineering & Applied Sciences · Depth: Expert, quick

Summary

An improved Fourier Neural Operator (FNO) is introduced for modeling two-dimensional Rayleigh-Bénard convection, demonstrating enhanced performance over a standard FNO baseline. This advancement is achieved by predicting time increments instead of full solutions, leading to superior accuracy. The proposed model is highly efficient, characterized by its compact size of 314k parameters and 1.26 MB, alongside a rapid 7 ms inference time. It successfully maintains accuracy levels comparable to those established in previous benchmarks. A critical observation from the research highlights that while FNOs exhibit generalization capabilities to finer meshes, their predictive accuracy ultimately remains limited by the inherent resolution of the training data utilized.

Key takeaway

For research scientists developing fluid dynamics models, this improved FNO offers a compelling alternative to standard approaches. If you are optimizing for both accuracy and computational efficiency in 2D Rayleigh-Bénard convection simulations, consider implementing time increment prediction. Your model could achieve higher accuracy with a compact 314k parameter footprint and 7 ms inference, but ensure your training data resolution is sufficient to avoid limiting overall performance.

Key insights

The improved FNO models Rayleigh-Bénard convection by predicting time increments, enhancing accuracy while remaining compact and fast.

Principles

• FNO accuracy is limited by training data resolution.
• Predicting increments improves FNO performance.
• FNOs generalize to finer meshes.

Method

The method involves an improved Fourier Neural Operator (FNO) that predicts time increments for 2D Rayleigh-Bénard convection, rather than full solutions, to achieve higher accuracy and efficiency.

In practice

• Use FNOs for efficient fluid dynamics simulation.
• Consider increment prediction for FNO accuracy.
• Evaluate training data resolution impact on FNOs.

Topics

• Fourier Neural Operators
• Rayleigh-Bénard Convection
• Fluid Dynamics Simulation
• Neural Operator Models
• Time Series Prediction
• Model Efficiency

Best for: AI Scientist, Research Scientist

Related on AIssential

• Self-Supervised Learning via Flow-Guided Neural Operator on Time-Series Data — FGNO enhances self-supervised learning for time-series data by dynamically varying corruption levels and using clean inputs for inference.
• Stability and Discretization Error of State Space Model Neural Operators — Theoretical bounds quantify discretization error and stability in neural operators for solving PDEs.
• Fourier Neural Operators with rank-1 lattice points and hyperbolic cross — Replacing FNO's tensor product grids with rank-1 lattices and hyperbolic cross improves efficiency and accuracy using fewer resources.
• SirenFNO: Efficient and Full Frequency Learning of Fourier Neural Operators — SirenFNO uses SIRENs and mode-wise kernel parameterization to enable full-frequency learning in FNOs, significantly reducing parameters.
• Multi-Rate Mixture of Experts for Accelerating Liquid Neural Network Training — Multi-Rate Mixture-of-Experts enhances Liquid Neural Networks for complex, multi-scale time-series modeling.
• Droplet-LNO: Physics-Informed Laplace Neural Operators for Accurate Prediction of Droplet Spreading Dynamics on Complex Surfaces — PI-LNO uses Laplace transforms and physics-informed neural networks to model droplet spreading dynamics efficiently.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Machine Learning.