AI weather models forecast atmospheric rivers too weak and too small, study finds

Atmospheric rivers drive both flooding and water supply, so forecast errors in their strength and size matter for the people who rely on predictions. The authors say the narrow, sharp-gradient structure of these storms may be poorly captured by the way machine learning models are trained.

The AI models also differed systematically from the traditional models in the biases of dynamic and thermodynamic fields inside strong atmospheric rivers. The authors say understanding these traits could help forecasters interpret AI predictions and improve both AI and traditional forecasts.

The researchers chose atmospheric rivers as a test case because these narrow, filament-like features have large gradients in temperature, moisture and wind fields. Such features may not be well represented under the loss functions used to train machine learning models.

The comparison covered medium-range forecasts out to 10 days for wintertime atmospheric rivers in the northern Pacific. It used both traditional gridpoint verification and new object-based methods that treat each river as a distinct feature, plus composites centred on the storms to diagnose physical behaviour.

Results were mixed. The AI models did well on some aggregate metrics, yet their rivers were slightly weaker and smaller than in reanalysis and traditional forecasts, and bias patterns inside strong events differed from those of GFS and IFS. The authors say this understanding can help forecasters interpret AI output and guide improvements to both model types.