Machine learning with non-Gaussian Data Assimilation
Abstract
The development of non-Gaussian and mixed distribution assimilation systems require an ability to detect a change in the distribution of the error to ensure that the associated data assimilation system is minimizing with respect to the most consistent distribution. Here we present results using machine learning techniques to switch a semi-static 3D VAR mixed Gaussian-lognormal cost-function between Gaussian and lognormal for the z component of the Lorenz 1963 model. Also as part of this work we shall present the reverse lognormal distribution that was detected as part of this work and present results when this third distribution is added as a further choice for the machine learning technique as well as the associated cost function, and properties of the distribution. Finally we have some plots to illustrate that these distributions are present in the atmosphere, and show that they change in times with what appear to different dynamical features.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2020
- Bibcode:
- 2020AGUFMNG0020007F
- Keywords:
-
- 3315 Data assimilation;
- ATMOSPHERIC PROCESSES;
- 3238 Prediction;
- MATHEMATICAL GEOPHYSICS;
- 3260 Inverse theory;
- MATHEMATICAL GEOPHYSICS;
- 3275 Uncertainty quantification;
- MATHEMATICAL GEOPHYSICS