Machine Learning Applications For Weather and Climate Modeling
Abstract
This study investigates the applications of machine learning (ML) to weather and climate modeling. We first show the potential for data-driven weather prediction by creating a low resolution, ML-based global atmospheric model that predicts the 3-dimensional atmosphere in the same for-mat as a physic-based numerical model. The ML-only atmospheric model is stable during 21-day forecasts and can reproduce large-scale atmospheric dynamics (e.g. Rossby waves). The ML-only model is able to outperform persistence and climatology for the first three forecast days in the midlatitudes. When compared to a simplified atmospheric general circulation model (AGCM), the ML-only model performs best for variables most heavily influenced by parameterizations in the AGCM (e.g. low level specific humidity).
Next, we combine a parallel, machine learning algorithm with a coarse resolution AGCM (SPEEDY) to create a hybrid atmospheric model. The hybrid model produces more accurate forecasts for all variables for at least the first 7 forecast days when compared to the host AGCM. Applications of the hybrid model for climate research are explored with a 11-year free run. The hybrid model is free of instability and can simulate the past climate with substantially smaller systematic errors and more realistic variability than the host AGCM.
Lastly, we show potential of ML for Earth System modeling by dynamically coupling a hybrid atmospheric model and a ML-based ocean model trained to predict the sea surface temperature (SST). The ML-only ocean model is able to reproduce SST dynamics with minimal biases for the past and present climate. The coupled model can simulate long-term variability in both the atmosphere and ocean (e.g. El Niño–Southern Oscillation). During a 70-year free run, we find that the coupled model does not exhibit climate drift and able to conserve total atmospheric mass and water vapor mass.
Citation
Arcomano, Troy J (2022). Machine Learning Applications For Weather and Climate Modeling. Doctoral dissertation, Texas A&M University. Available electronically from https : / /hdl .handle .net /1969 .1 /198600.