This repository is the code associated with the WAF manuscript titled: "A Machine Learning Tutorial for Operational Meteorology, Part II: Neural Networks" written by Chase, R. J., Harrison, D. R., Lackmann, G. and McGovern, A. in prep. While the paper undergoes review, feel free to read its preprint and provide any comments via email to the corresponding author. The SSEC-hosted branch of this repository has been edited for use in a 2026 AI/ML short course, and any problems encountered here are not the responsibility of the original creators with AI2ES. If you have any issues with the code (bugs or other questions) please leave an issue associated with this repo.
This second paper and repo (of two) that covers neural networks and deep learning methods (if you don't know what these phrases even mean thats OK! Check out Section 2 in the paper). If you dont know what machine learning is, please check out paper one in this series found here.
Note, this figure will be updated in 2023 to contain the full results of 2022.
We showed this motivation in the previous part of the paper. Now we would like to draw your attention to subplot (c) where the number of meteorology papers with neural network methods are shown. This group of meteorology papers is growing rapidly. Thus, we have written this part of the tutorial to hopefully help meteorologists gain some intution and perspective about neural networks.
Like part 1, we go beyond just discussing neural networks in an abstract way. We again use the The Storm EVent ImagRy (SEVIR) dataset
Unfortunately the original SEVIR dataset is far to combersome for most to deal with on their personal machines (the size of the data is almost 1 TB). So to assist in making an accessible and tangible sandbox dataset for this tutorial we made sub-SEVIR, which decreases the resolution of all variables to be 48 by 48 (~8km by 8km) and only take 1 hour of data instead of all 4 hours in the SEVIR dataset. The result is a much more magangable dataset (only 2 GB in storage now). So you can see the change in the dataset,here is a movie from the original SEVIR dataset and resolution for an example storm. The sub-SEVIR version can be seen here.
Unfortunately 2 GB is still too big for github, so we have made sub-SEVIR available by download on zenodo 
There are two main ways to interact with the code here.
This is the recommended and the quickest way to get started and only requires a (free) google account. Google Colab is a cloud instance of python that is run from your favorite web browser (although works best in Chrome). If you wish to use these notebooks, navigate to the directory named colab_notebooks.
Once in that directory, select the notebook you would like to run. There will be a button that looks like this once it loads:
Click that button and it will take you to Google Colab where you can run the notebook. Please note it does not save things by default, so if you would like to save your own copy, you will need to go to File -> save a copy in drive
Google colab is awesome for those who do not know how to install python, or just dont have the RAM/HDD locally to do things. You can think of it this way. This notebook is just providing the instructions (i.e., code) to do what you want it to. Meanwhile the data and physical computer are on some Google machine somewhere, which will execute the code in the notebook. By default this google owned machine will have 12 GB of RAM and about 100 GB of HDD (i.e. storage).