The bottleneck holding back sharper AI weather forecasts turns out to be data, not algorithms.
A new research framework called BaguanHR tackles a specific problem: machine-learning models that forecast weather at a fine 0.1-degree resolution have almost nothing to train on, since decades of reanalysis records exist only at the coarser 0.25-degree scale. Instead of the usual fix, fine-tuning a 0.25-degree model on whatever scraps of 0.1-degree data exist, the researchers found that approach loses information that can't be recovered. Their workaround is to use super-resolution to synthesize large volumes of 0.1-degree data from the existing 0.25-degree ERA5 archive, then train on that synthetic-plus-real mix. The resulting model beat both other ML forecasting systems and IFS-HRES, the physics-based benchmark, across more than 85% of lead times within a 72-hour window.
This matters because the field has largely treated better forecasts as an architecture problem, chasing bigger or cleverer models. This paper argues the ceiling is really about training data availability, and that super-resolution is a cheap way to manufacture more of it. The team also found a power-law pattern: doubling the training data cut forecast error (RMSE) by 4.6% at 72 hours and 4.9% at 120 hours.
Those are real but modest gains for a data doubling, not a leap. And synthetic high-resolution data, however cleverly generated, is still a proxy for the real thing IFS-HRES was built on. Worth watching, not worth the confetti.