Hannah Christensen (she/her)

Abstract:

Stochastic parameterizations account for uncertainty in the representation of unresolved subgrid processes by sampling from the distribution of possible subgrid forcings. Some existing stochastic parameterizations utilize data‐driven approaches to characterize uncertainty, but these approaches require significant structural assumptions that can limit their scalability. Machine learning models, including neural networks, are able to represent a wide range of distributions and build optimized mappings between a large number of inputs and subgrid forcings. Recent research on machine learning parameterizations has focused only on deterministic parameterizations. In this study, we develop a stochastic parameterization using the generative adversarial network (GAN) machine learning framework. The GAN stochastic parameterization is trained and evaluated on output from the Lorenz '96 model, which is a common baseline model for evaluating both parameterization and data assimilation techniques. We evaluate different ways of characterizing the input noise for the model and perform model runs with the GAN parameterization at weather and climate time scales. Some of the GAN configurations perform better than a baseline bespoke parameterization at both time scales, and the networks closely reproduce the spatiotemporal correlations and regimes of the Lorenz '96 system. We also find that, in general, those models which produce skillful forecasts are also associated with the best climate simulations.

Abstract:

The impact of a warming climate on El Nino-Southern Oscillation (ENSO) is investigated in large ensemble simulations of the Community Earth System Model (CESM1). These simulations are forced by historical emissions for the past and the RCP8.5-scenario emissions for future projections. The simulated variance of the Nino-3.4 ENSO index increases from 1.4◦C2 in 1921-1980 to 1.9◦C2 in 1981-2040 and 2.2◦C2 in 2041-2100. The autocorrelation timescale of the index also increases, consistent with a narrowing of its spectral peak in the 3- to 7-yr ENSO band, raising the possibility of greater seasonal to interannual predictability in the future. Low-order linear inverse models (LIMs) fitted separately to the three 60-yr periods capture the CESM1 increase in ENSO variance and regularity. Remarkably, most of the increase can be attributed to the increase in the 23-month damping timescale of a single damped oscillatoryENSO eigenmode of these LIMs by 5 months in 1981-2040 and 6 months in 2041-2100. These apparently robust projected increases may however be compromised by CESM1 biases in ENSO amplitude and damping timescale. A LIM fitted to the 1921-1980 observations has an ENSO eigenmode with a much shorter 8-month damping timescale, similar to that of several other eigenmodes. When the mode’s damping timescale is increased by 5 and 6 months in this observational LIM, a much smaller increase of ENSO variance is obtained than in the CESM1 projections. This may be because ENSO is not as dominated by a single ENSO eigenmode in reality as it is in the CESM1.