Ensuring Stability in Data-driven Climate Modeling: Insights from a 1D Gravity Wave-QBO Testbed

Hamid Pahlavan

NorthWest Research Associates

Tuesday, Oct 08, 2024, 2:00 pm MT
DSRC Room GC402

JOIN THE WEBINAR:

Abstract

Machine learning (ML) techniques, especially neural networks (NNs), have shown promise in learning subgrid-scale parameterizations for climate models. However, a major problem with data-driven parameterizations, particularly those learned with supervised (offline) algorithms, is model instability. Current remedies are often ad-hoc and lack a theoretical foundation.

Here, we combine ML theory and climate physics to address a source of instability in NN-based parameterization. We demonstrate the importance of learning spatially non-local dynamics using a 1D model of the quasi-biennial oscillation (QBO) with gravity wave (GW) parameterization as a testbed. While common offline metrics fail to identify shortcomings in learning non-local dynamics, we show that the receptive field (RF) can identify instability a-priori.

We find that NN-based parameterizations that seem to accurately predict GW forcings from wind profiles (R2≈0.99) cause unstable simulations when RF is too small to capture the non-local dynamics, while NNs with the same number of parameters but large-enough RF are stable. We examine three broad classes of architectures—convolutional NNs, Fourier neural operators, and fully-connected NNs—with the latter two inherently possessing large RFs.

Furthermore, we demonstrate that learning non-local dynamics is crucial for ensuring the stability and accuracy of a spatiotemporal emulator of the zonal wind field. Given the ubiquity of non-local dynamics in the climate system, we expect the use of effective RF, which can be computed for any NN architecture, to be important for many applications.

Additionally, we explore retraining strategies for NNs initially trained offline in a small-data regime, which yield unstable QBOs once coupled to the 1D model. We demonstrate that online re-training of just two layers of this NN using ensemble Kalman inversion and only time-averaged QBO statistics (e.g., period and amplitude) leads to parameterizations that yield realistic QBOs. Fourier analysis of the NNs' kernels suggests why/how re-training works and reveals that these NNs primarily learn low-pass, high-pass, and a combination of band-pass filters, potentially related to the local and non-local dynamics in GW propagation and dissipation.

These insights underscore the necessity of integrating ML theory with climate physics to design and analyze data-driven algorithms for weather and climate modeling.

Bio: I am an atmospheric scientist currently holding a joint position as a Research Scientist at Northwest Research Associates (NWRA), Boulder, and the University of Chicago, where I work closely with Dr. M. Joan Alexander and Prof. Pedram Hassanzadeh. My research leverages advances in machine learning (ML) alongside the principles of climate dynamics to improve climate modeling and analysis. My recent work addresses critical challenges in AI-driven climate modeling, particularly issues of instability and the scarcity of high-fidelity data.

I completed my Ph.D. in Atmospheric Sciences at the University of Washington in 2022 under the guidance of Profs. John M. Wallace and Qiang Fu. My doctoral research focused on the dynamics of the quasi-biennial oscillation (QBO), the dominant mode of interannual variability in the tropical stratosphere, and on atmospheric gravity waves (GWs). This work contributed to a deeper understanding of stratospheric processes, which play a critical role in modulating weather and climate variability and have significant implications for extreme events such as cold air outbreaks and heat waves—both of which pose serious risks to public health and infrastructure.