Peering into the black box of machine learning parameterization
Abstract
Machine learning has recently been applied to improve the moist physics parametrizations used by atmospheric models, but many challenges remain unsolved. For instance, a neural network (NN) trained to mimic the moist physics of a near-global aqua-planet simulation with a 4 km resolution achieves excellent diagnostic performance but produces numerical instability when coupled to a general circulation model with a 160 km resolution. Weather forecasts with this setup crash after 5-9 days. The linearized response of this NN to its inputs is computed by automatic differentiation and shows that the NN is sensitive to physically-implausible inputs. Training the NN without these inputs stabilizes the simulation, which accurately predicts the "weather" of the training data. Coupling both stable and unstable versions of the NN to a linear model for atmospheric gravity waves hints at the instability mechanisms of these NNs. The eigenvalues of this linear system correspond to the wave speeds and growth rates of atmospheric waves. The eigenvectors encode the vertical velocity, temperature, and humidity structures of these waves. The stable NN produces a physically-plausible moisture-mode with an e-folding time of .8 days, but most other modes are strongly damped. On the other hand, the unstable NN produces many modes with physically-implausible structures, which could cause numerical instability in coupled simulations. Overall, this linearized analysis reveals coupled instabilities at much less computational expense than a full nonlinear simulation and could provide new insights into tropical meteorology.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2019
- Bibcode:
- 2019AGUFMIN44A..03B
- Keywords:
-
- 3365 Subgrid-scale (SGS) parameterization;
- ATMOSPHERIC PROCESSES;
- 1910 Data assimilation;
- integration and fusion;
- INFORMATICS;
- 1920 Emerging informatics technologies;
- INFORMATICS;
- 1942 Machine learning;
- INFORMATICS