Weather downscaling no longer just tunes the backbone: PROLA splits a fixed rank budget between principal and orthogonal directions, cutting Z500 error to 50.58
Synopsis
Mapping coarse-resolution history directly to future high-resolution weather trajectories requires forecasting, spatial downscaling and temporal densification at once, yet existing low-rank fine-tuning does not separate reusing pretrained principal directions from correcting outside them under a fixed rank budget. PROLA splits that budget into a principal branch and an orthogonal branch and rescales rank-1 optimizer updates with gradient signals; adapting Aurora on WeatherBench 2 with 3.92M trainable parameters, it reaches a 3-hour Z500 RMSE of 50.58, below AdaLoRA's 53.40.
Interpretation
A weather foundation model can be adapted to emit future fields on a finer spatial grid and a denser time step without full fine-tuning for every target resolution. Earlier weather forecasting kept the spatial and temporal grids fixed, spatial downscaling refined the grid at a single time, and temporal downscaling densified a trajectory on the same grid, with the three usually handled separately. On WeatherBench 2 with Aurora as the pretrained backbone, PROLA attains the lowest RMSE and highest ACC among the compared low-rank methods on all five representative variables at both the 3-hour and 6-hour leads.
The adaptation splits prediction into a spatial stage and a temporal stage: the backbone first produces a high-resolution field at the final lead time, and the trajectory is built from it. PROLA is a low-rank parameter-efficient fine-tuning method that divides a fixed rank budget between a principal branch initialized from the dominant singular directions of the pretrained weights and an orthogonal branch projected outside those directions. At a fixed total rank of 8, the balanced 4-and-4 split gives a 3-hour Z500 RMSE of 51.36, below 53.66 for the principal branch alone and 54.54 for the orthogonal branch alone; adding gradient-guided update scaling lowers it further to 50.30.
Fixed numerical scaffolds carry the large-scale structure while lightweight residual heads supply detail that interpolation cannot recover. The spatial stage lifts coarse inputs to the target grid by bilinear interpolation, the temporal stage interpolates linearly between the upsampled present state and the predicted endpoint, and two correction heads learn the residuals. Removing both correction heads raises the 3-hour Z500 RMSE to 87.73, showing the scaffolds alone are insufficient; removing only the spatial head gives 84.53 and removing only the temporal head gives 53.35.
Perspective
The evaluation covers the WeatherBench 2 global reanalysis benchmark, ERA5-to-CERRA regional downscaling over Central Europe and regional precipitation forecasting; the main comparison uses the Aurora backbone and reports means and standard deviations over three seeds.
The authors note that the task is constrained by atmospheric chaos and the limited fine-scale information in coarse inputs, and list broader regions and target sources, physical constraints and probabilistic prediction as future directions.