AdaptFNO: Adaptive Fourier Neural Operator with Dynamic Spectral Modes and Multiscale Learning for Climate Modeling

Fourier Neural Operators (FNOs) are effective for modeling spatio-temporal dynamics but often favor low-frequency patterns, overlooking fine-scale details critical in climate forecasting. We propose AdaptFNO, an adaptive variant that dynamically adjusts spectral modes based on input frequency content and partitions domains into frequency-specific patches. A cross-attention mechanism aligns global and local features, enabling efficient multiscale learning. Evaluated on ERA5 reanalysis data, AdaptFNO captures both large-scale circulation and fine-grained events, such as typhoon trajectories, while maintaining long-range stability. Preliminary results on Typhoon Yagi highlight its ability to preserve details of cyclone formation, showing promise for high-resolution climate forecasting. The source code is available at: https://github.com/HySonLab/AdaptFNO