Analog ensemble data assimilation with machine-learning methods to construct a forecast analog ensemble
Abstract
Ensemble forecasting is a computational bottleneck for data assimilation using coupled Earth system models or their high-resolution components. "Analog ensemble data assimilation" runs one or a few model forecasts, and then creates an ensemble of analogs of the forecast for use in ensemble data assimilation. From a Bayesian perspective, the analog ensemble is drawn from an "analog prior", which can be defined in many ways. One option involves a catalog of model states, from which analog ensemble members are selected based on their similarity to the model forecast. Another option involves training a generative machine-learning model to construct analogs of the model forecast. This presentation presents analog ensemble data assimilation using both "found" analogs and analogs constructed using autoencoders and variational autoencoders in the context of a multiscale Lorenz-96 model with 2,624 variables. Preliminary results from a coupled atmosphere-ocean model with nearly ten million variables are also presented. Despite promising results for such a high-dimensional model, it remains unlikely that one could train a generative model to construct complete states of a high-resolution or coupled Earth system model. A "patching" strategy is therefore proposed to construct analogs by stitching together reconstructions of subdomains ("patches"); the patched version of the method is shown to outperform the original method using global analogs in the multiscale Lorenz-96 model.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2021
- Bibcode:
- 2021AGUFMNG24A..01G