Please use this identifier to cite or link to this item:
http://arks.princeton.edu/ark:/88435/dsp01wm117s18c
Title: | Physics-Informed Neural Networks for Tropical Cyclone Data Assimilation and Modeling |
Authors: | Eusebi, Ryan |
Advisors: | Vecchi, Gabriel Lai, Ching-Yao |
Department: | Computer Science |
Class Year: | 2022 |
Abstract: | Many studies to date have demonstrated the importance of proper initial conditions for hurricane forecast models. However, imperfect initialization schemes still contribute to forecast errors, especially in intensity forecasts which have only seen modest improvements in recent times. We propose a physics-informed neural network (PINN) for the task of initializing tropical cyclone (TC) flow fields and assimilating observational data into the fields. We investigate the PINN's ability to not only reconstruct the flow field for the purpose of initialization, but to model the evolution of the flow using sparse observations recorded within the storm. To assess the model, we use output from the Geophysical Fluid Dynamics Laboratory's forecast model SHiELD's forecast of category 4 Hurricane Ida (2021). We sample synthetic flight paths through the storm and use the PINN to reconstruct the 2-Dimensional TC flow field solely based on the observations sampled along these transects through the eye of the storm. We find that using minimal and sparse data, the PINN is able to accurtaely reconstruct the full wind and pressure fields and interpolate and extrapolate in between and beyond times at which observations are provided. It achieves root mean square errors (RMSE) of 1.5-2 m/s averaged over the full TC grid, with RMSEs of 3-6 m/s averaged over the eyewall. RMSE values are generally 7-10% of the target wind speed values. We find the PINN accurately captures the radial wind profiles of the storm at various points in the forecasts, and the outputted profiles are physically consistent with the governing equations provided to the PINN. The results highlight the power of PINNs at encoding partial differential equations (PDE) and assimilating data to find PDE solutions. We also show that PINNs offer a much more interpretable model than other "black box" machine learning methods. |
URI: | http://arks.princeton.edu/ark:/88435/dsp01wm117s18c |
Type of Material: | Princeton University Senior Theses |
Language: | en |
Appears in Collections: | Computer Science, 1987-2023 |
Files in This Item:
File | Description | Size | Format | |
---|---|---|---|---|
EUSEBI-RYAN-THESIS.pdf | 7.29 MB | Adobe PDF | Request a copy |
Items in Dataspace are protected by copyright, with all rights reserved, unless otherwise indicated.