Skip to content


Machine Learning for Data Assimilation

Accurate and timely observations are critical for making good weather forecasts, and data assimilation is the process that brings observations into weather models. Machine learning provides new opportunities to improve and extend data assimilation capabilities. For example, convolutional neural networks can make use of spatial context and texture in satellite images, which is currently going unused by weather forecasts. CIRA is performing the research to bring ML capabilities to operational data assimilation. Figure 1 (adapted from Back et al. 2021) provides an example of this research. The Rapid Refresh Forecast System (RRFS) uses radar-based estimates of the latent heating from clouds and precipitation (Fig. 1A) to initialize convection in short-range forecasts (Fig. 1B). This provides much better forecasts than using no radar initialization (Fig. 1C). At CIRA, we developed the GREMLIN machine learning model (Hilburn et al. 2021) that translates GOES-R radiances and lightning into synthetic radar reflectivity fields. Estimates of latent heating from GREMLIN (Fig. 1D) were used to initialize forecasts (Fig. 1E) that produce convection better matching observations (Fig. 1F). Additional analysis from this one-week retrospective simulation showed that using GREMLIN to initialize convection produced more skillful forecasts out to lead times of 12 hours over the western U.S. where radar coverage is spotty due to terrain blockage. These promising results show that machine learning offers new opportunities to use satellite observations to improve weather forecasts of high-impact events.

Figure 1. RRFS retrospective simulation evaluating GREMLIN, results from Back et al. (2021). (A) MRMS radar-based latent heating temperature tendency currently used in RRFS, (B) 1-hour forecast composite radar reflectivity using the radar-based latent heating, (C) 1-hour forecast


Hilburn, K. A., Ebert-Uphoff, I., and Miller, S. D., Development and Interpretation of a Neural Network-Based Synthetic Radar Reflectivity Estimator Using GOES-R Satellite Observations. Journal of Applied Meteorology and Climatology,, Jan 2021.

Back, A., Weygandt, S., Alexander, C., Benjamin, S., Hu, M., Ge, G., James, E., Kliewer, A., Mecikalski, J., Dowell, D., Bruning, E., Hilburn, K., and Sebok, A., Convection-indicating GOES-R products assimilated in the experimental UFS Rapid Refresh System. AGU Fall Meeting, Presentation A22B-02, 14 Dec 2021.

Primary contact:

Kyle Hilburn