Variance predictions of atmospheric variables
The steps in this guide should be followed in order.
Copy, edit and rename data_prep/python/script_variables_template.py to data_prep/python/script_variables.py such that the variables satisfy the needs of your setup. This must include root directories for reading and storing the data.
Ensure python3 is ready and running the following command:
python loop_files.py <day>
where <day> is an integer. In this case, day takes values 1, 2, ..., 10.
run_loop_files.sh is already setup for run this.
Run the script /speedy/model/run/run_first.sh. This builds the SPEEDY executable and simulates a 12 month forecast: 1981/01/01-1982/01/01.
This script takes about 5 minutes to complete.
Run the script /speedy/model/run/run_cylce.sh. This script rebuilds the SPEEDY executable to run in 6 hourly chunks. This is necessary for our script later. run_cycle.sh also simulates a 10 year forecast which starts by using the final output from run_first.sh.
This script takes about ?? to complete
Copy, edit and rename hybrid/src/script_variables_template.py to hybrid/src/script_variables.py such that the variables satisfy the needs of your setup. This must include root directories for reading and storing the data.
This file includes a flag for using a pre-trained GP model or training a new model when before launching SPEEDY. This default is to True i.e. train a new GP everytime.
The script /hybrid/src/train-gp.sh provides an example of a submission script for training the Guassian process. This example is designed for submission UCL's Myriad HPC.
At the time of writing, the latest verion of the python package mogp-emulator, version 0.7.2, does not work for our needs. Instead, we revert to package version 0.6.1 and install directly from the GitHub repository.
Now turn the TRAIN_GP flag to False in your script_variables.py file. Future steps will load in the pre-trained Gaussian process.
The script /hybrid/src/run-speedy.sh launches the wrapper.py script which will load the pre-trained Gaussian process and run SPEEDY. Each time step within the main loop will relaunch SPEEDY to run for a 6 hour forecast. This loop will produce 10 years worth of forecast where each 6 hour simulation represents a different potential atmosphere as sampled from the Gaussian process.
A few changes have been made to this code to suit our scientific needs.
Line 16 of /speedy/model/source/makefile has been commented out to ensure endianness of the programm remains consistent with the system default.
See analysis/READ_ME.md for details on running the scripts which generate the figures in the manuscript.
To download the ground truth ERA5 and GPCP data see data_download directory. The Global Precipitation Climatology Project data can be downloaded from Copernicus but please ensure that you have set up the api correctly. Instructions can be found here. The ERA5 temperature and specific humidity data is downloaded from the WeatherBench2 project.