Fig. 14: 4DVarFormer for 3D-multivariate weather state estimation. | npj Climate and Atmospheric Science

Fig. 14: 4DVarFormer for 3D-multivariate weather state estimation.

From: Accurate initial field estimation for weather forecasting with a variational constrained neural network

Fig. 14

a Architecture of 4DVarFormer, a 4DVar-constrained Transformer model. The model takes the background field (xb(t0)) and observations (y(t0), y(t1), ⋯ , y(t3)) as input and outputs the analysis increment (Δx(t0)). Then, the increment is added to the background field to get the final analysis field. b 4DVar Gradient Block, a neural implementation of the gradient of the 4DVar cost function based on a pre-trained FourCastNet. c Assimilation Network, a neural network that directly provides the analysis increment Δx(t0).

Back to article page