The Next Generation
Extreme rainfall from tropical cyclones (TCs) is an increasing hazard. Improvements in the accuracy of prediction of rainfall rates in space and time are critical for designing and planning for climate adaptation. TCs reflect a large-scale organization of the atmosphere and hence so does the ensuing rainfall. We propose to go beyond modeling TC tracks and winds to develop a physics-informed Machine Learning approach to parameterize the spatio-temporal rainfall fields. The approach could be used to enhance the modeling of TC precipitation at sub-season to season for a decade or longer using physics-based model projections of climate dynamics. Ambient conditions in which tropical cyclones are embedded will be analyzed with TC track data, high-resolution 3-hourly precipitation, daily and monthly historical data from climate reanalysis models. Linking convolutional neural nets, Gaussian processes, and functional data analysis we propose to simulate TC tracks and the associated rainfall conditional on the ambient conditions relevant for each physical process. This approach leverages the strength of the physics-based models in simulating the atmospheric circulation fields and variables that determine the TC tracks and precipitation while overcoming their weakness in representing the tracks and associated precipitation dynamics. It addresses the nonstationarity in the risk profile of flooding induced by TCs. The results would aid infrastructure design, insurance pricing, and disaster response planning for communities that cannot currently access this information for future conditions from the next season to a decade and longer.
We are now ready to develop the next generation of tropical cyclone track and precipitation model building on our experience by appropriately integrating physics into the evolution of the TCs and their precipitation fields and using a wider set of machine learning tools. As in the prior work, we consider two aspects: (1) TC birth and track generation, and (2) TC related precipitation fields over land. In each case, consider that ambient conditions are available as spatial fields of the appropriate variables of interest from a climate model (reanalysis or seasonal forecast or climate change projection). The challenge is to functionally relate the TC attribute in each of these steps to the spatial fields of ambient conditions.
For TC path prediction, we can represent each TC track as a smooth curve that runs through its observed positions every three hours. Deep learning models, like convolutional neural networks or similar tools, can be used to analyze vast maps of environmental variables—helping update and refine the predicted path as the TC evolves. These models are designed to focus only on the most important features, which improves prediction accuracy and reduces computation.
To estimate rainfall, a second model would be used. This model would predict how much rain will fall near the TC’s current position by combining statistical methods and deep learning. It would capture both typical rainfall patterns and the uncertainty that comes with chaotic weather systems. All models would be first trained and then validated using actual observations, and once proven reliable, they could be applied to test different climate scenarios.
In essence, by blending flexible mathematical modeling, deep learning, climate physics, and statistical analysis, this proposed approach aims to provide more accurate forecasts for TC tracks and rainfall, both in current conditions and under future climate changes.