Are AI Models Capturing Atmosphere Dynamics?
12.810 Final Proj
Ken Yan
Cover photo generated by ChatGPT
NWP vs AI Weather Forecast
FourCastNet: Fourier ForeCasting Neural Network
Spherical Fourier Neural Operator
FourCastNet (NVIDIA) Pangu (Huawei)
Literature Review
“Yes!” Papers
“Not there yet” Papers
Data and Methods
AI models and Gradient Wind Balance
AI Model Outputs Availability
Gradient Wind Balance (Sun et al., 2024)
Choose TC from 2018-2024 (testing set)
Weak Phase init, model = pangu
Choose TC from 2018-2024 (testing set)
Strong Phase init, model = fourcastnetv2
Compared Quantity: Extracted Wind vs Gradient Wind
Hurricane Lee, init = 2023/09/05 (Weak Phase), lead = 144 hour
Hurricane Lee, init = 2023/09/05 (Weak Phase), lead = 144 hour
6 cat5 hurricanes considered!
Beryl (2024 June)
Dorian (2019 Aug)
Lorenzo (2019 Sep)
Ian (2022 Sep)
Lee (2023 Sep)
Milton (2024 Oct)
Weak Phase init
Strong Phase init
“…The lack of the expected balance, i.e., the difference between the full wind and the gradient wind, is most noticeable in forecasts initiated during the strong phase of the TCs.” (Sun et al., 2024)
Supp Slides
Typhoon Bebinca (upper right)
Dessert Bebinca (lower right)