3DVar for Neural-Network based observation operator

This repository contains the implementation of a three-dimensional variational (3DVar) data assimilation system designed to work with the nonlinear neural-network-based observation operator (HNN) that maps ALADIN model outputs to radar–reflectivity observation space. The codebase includes several 3DVar configurations and auxiliary datasets needed to run and test the system. The directory provides:

1. A complete 3DVar implementation using a diagonal background error covariance matrix B combined with a recursive filter.
2. A 3DVar variant that uses only the HNN observation operator to compute analysis increments.
3. A 3DVar setup for single-observation assimilation experiments, useful for testing, diagnostics, and sensitivity studies.
4. A subdirectory containing precomputed B-matrix datasets used to characterise background error statistics.
5. A subdirectory containing the observation-space mask defining where radar–reflectivity observations are assimilated.

Together, these components provide a flexible framework for experimenting with neural-network-based observation operators in a variational data assimilation context.

To use it, you need:
1 - The ALADIN model and radar observation datasets: LISCA-ALADIN HNN (doi: 10.5281/zenodo.17880622)

2- The HNN operator: 3DVar Neural Network-Based Observation Operator (doi: 10.5281/zenodo.17898084)