Initializing a Numerical Weather Model with Neural Network Data Assimilation

Data assimilation is crucial in numerical weather prediction (NWP). Techniques such as variational methods and Kalman Filter-based approaches, though varied in their implementation, aim to combine atmospheric observations from sources like satellites and surface stations with a background (B), an a priori condition, to create an initial state that closely mirrors the actual atmosphere. This process significantly improves weather forecast accuracy.

In our study, we developed a model using Convolutional Neural Networks (CNNs) within a generator framework that integrates upper-air, surface, and satellite data (AMSU-A, HIRS-4, and a global combination of geostationary satellites). The goal is to estimate the referenced analysis (RA), which in this case is based on ERA5 reanalysis. To generate the RA, we combined 24 to 48-hour operational forecasts from the ECMWF center with surface observations and infrared geostationary composite data.

For the initial performance assessment, we tested the model on a case outside the training set, using a 24-hour GFS forecast as the background (B). We estimated variables including Temperature, Geopotential, Wind U Component, Wind V Component, and Relative Humidity across 13 vertical levels, with a horizontal resolution of 0.25x0.25 degrees. The MPAS (Model for Prediction Across Scales) numerical weather prediction model was then initialized with an approximate resolution of 15 km.

Preliminary results show that initializing the system with the Neural DA system reduced surface forecast errors in the MPAS model for up to 10 days compared to using the GFS forecast. We analyzed surface data from the GTS system, focusing on surface variables and parameterizations that were consistent between the two models. Additionally, surface pressure tendencies shows similar stability between the Neural DA and GFS-initialized systems

The next step in this research is to test additional cases, particularly specific weather events in Brazil. We will compare the performance of the neural network-generated analyses with those from major centers like NCEP and ECMWF for the same periods to evaluate the model’s broader applicability.