Earth systems modeling is a complex and computationally expensive task, especially when it comes to predicting the intensity of extreme weather events like tropical cyclones. Traditional methods often fall short in terms of speed and accuracy due to the sheer scale and complexity of current models. Enter the realm of Machine Learning (ML), for earth systems modeling.

Our team at GDA recently released a publication in which we apply ML methodologies to improve the speed and accuracy of hurricane modeling. By leveraging data from in situ observations and advanced computational techniques, we demonstrate how ML can automate analysis and model development, making earth systems modeling more efficient and effective.

Hurricane Modeling

Tropical cyclones are extreme water cycle events which can have a devastating impact on the environment, economy, and human lives. Modern weather methodologies have improved hurricane path prediction, but intensity prediction remains a challenge. To address this, we focused on assimilating data at the air-sea interface, a crucial area for understanding storm intensity.

Sensor-laden floats were deployed in the path of Hurricane Sally (Sept. 2020) as part of the DARPA Ocean of Things (OoT) program. These floats gathered several modalities of environmental data. By integrating this data into a traditional hurricane model, our team gained insights into the storm’s intensity profile and the interactions between ocean currents, wave dynamics, and atmospheric effects.

Drifters and weather stations in the Gulf of Mexico with Hurricane Sally.

ML in Action

ML techniques, specifically deep neural networks (DNNs), can be employed to correct these models and address discrepancies between theoretical predictions and observed data. This approach, also applied in turbulence modeling, offers novel ways to assimilate data and improve model accuracy.

Beyond correcting models, DNNs like the one employed here can also be used to recover critical parameters from data, such as the relationship between sensor float motion and wind and water forces. DNNs excel in this area due to their speed and ability to generalize from data.

The image below depicts the difference between a standard, circular model of Hurricane Sally’s pressure gradient and the pressure gradient model augmented with our DNN-informed parameters and corrections.

Isobars of pressure from a circularly symmetric model vs a model augmented via DNN

Conclusion

The use of ML in earth systems modeling provides new possibilities in how we study and predict extreme weather events. Researchers can now achieve faster and more accurate predictions by combining domain expertise with robust ML models. This enhances our ability to respond to and mitigate the impacts of these events.