Separating GraphCast Model Bias from ERA5 Reanalysis Error

Determine the relative contributions of GraphCast model-specific bias and ERA5 reanalysis error to the forecast improvements achieved by gradient-based optimization of initial conditions in GraphCast, given that GraphCast is trained on ERA5 and the study verifies forecasts against ERA5.

Background

The paper uses GraphCast, a machine-learning weather model trained on ERA5, to optimize initial conditions via gradient descent and demonstrates substantial forecast error reductions and extended predictability horizons. Cross-model validation with Pangu-Weather shows smaller but significant improvements, suggesting that the optimized initial conditions encode a mix of analysis corrections and model-specific bias mitigation.

Because GraphCast is trained on ERA5 and the verification is performed against ERA5, disentangling how much of the improvement arises from correcting ERA5 analysis errors versus compensating for GraphCast’s model bias is nontrivial. The authors explicitly state that this separation remains ambiguous, highlighting a key unresolved question about interpreting the source of the observed forecast gains.