Fourier Neural Operators for Rayleigh-Bénard Convection

This paper presents an improved Fourier Neural Operator (FNO) approach for modeling two-dimensional Rayleigh-Bénard convection in computational fluid dynamics. The key innovation lies in predicting time increments rather than full solutions, which enables higher accuracy compared to standard FNO baselines. The resulting model achieves a compact architecture with 314k parameters (1.26 MB) and performs inference in just 7 milliseconds. The authors demonstrate that while FNOs can generalize to finer meshes, the accuracy remains fundamentally constrained by the resolution of the training data, providing important insights for neural operator applications in thermal convection problems.