3DVar Neural Network-Based Observation Operator

This repository provides the implementation of a nonlinear observation operator H for mapping ALADIN numerical weather prediction model outputs to radar–reflectivity observation space. The operator is implemented as a convolutional encoder–decoder neural network based on a residual U-Net (ResUNet). The network takes multi-level ALADIN model fields as input and produces radar reflectivity at the same horizontal resolution. The directory structure includes a TEST subdirectory containing a ready-to-use trained ResUNet model. The dataset necessary to train the model is: LISCA-ALADIN HNN (doi: 10.5281/zenodo.17880622).