Calibrating parameterizations with ensemble Kalman methods and high-resolution data
Abstract
The remarkable achievements of machine learning (ML) over the last decade have led to renewed interest in data-driven methods for Earth system modeling. The spotlight is often on improving parametric models of physical processes that are associated with uncertainty in climate projections. Among these, fluid transport processes can be accurately simulated in small domains, which can provide an invaluable amount of high-resolution data. Data-assimilation methods that provide a cheap and flexible way to ingest these data are a crucial step in the development of online learning strategies for climate models. In this context, we propose parameter calibration strategies using high-resolution data that are highly parallelizable, gradient-free, and require no domain-specific knowledge. These characteristics make the proposed methods particularly effective for the calibration of computationally expensive or stochastic parametric models. Furthermore, their reliance on ensemble Kalman methods make the strategies amenable to combination with ML-based uncertainty quantification. The strategies are demonstrated through the calibration of a unified turbulence and convection model.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2021
- Bibcode:
- 2021AGUFM.A55C1387L