search

google-research / neuralgcm

Hybrid ML + physics model of the Earth's atmosphere
https://neuralgcm.readthedocs.io
Apache License 2.0
616 stars 65 forks source link

Inaccuracies in Atmospheric Forecasts at High Altitudes, Particularly Geopotential Height (gh) #123

Open leanderloew opened 6 days ago

leanderloew commented 6 days ago

While running the example code on Colab using L4 and T4 GPUs, I observed inaccuracies in the forecasts high in the atmosphere. Specifically, geopotential height (gh) exhibits unexpected surface features at elevated atmospheric levels, which seems very strange.

This is for the neural_gcm_dynamic_forcing_stochastic_1_4_deg.

Screenshot for Reference:

Screenshot 2024-09-13 at 15 25 00
kochkov92 commented 3 days ago

Hi @leanderloew - thank you so much for documenting this! This is in part an expected behavior of our current models - we do not optimize predictions on pressure levels in 0-30 hPa range (even though model ingests all 37 pressure levels from ERA5). The rationale for this choice is that with 32 equidistant sigma levels the top of the atmosphere cannot be resolved well in the model representation.

Unfortunately this remark got somewhat lost in the paper, currently only briefly mentioned in appendix I1 3rd paragraph.