LPDM-Sensitivity-Emulation

This repository contains a number scripts and links to datasets required to reproduce results presented in "Emulation of greenhouse-gas sensitivities using variational autoencoders", by Laura Cartwright, Andrew Zammit-Mangion, and Nicholas M. Deutscher.

All scripts (except one shell script) are written in R, specifically, R 3.6.3. The easiest way to run them (and the way described below) is via the terminal/command line, and using the Conda environment included in this repository. The scripts can also easily be run via RStudio. Note you must download and install R before RStudio.

Downloading the repository

This can either be done by clicking the green "Clone or download" button, or by running the following in the command line/terminal:

git clone https://github.com/Lcartwright94/LPDM-Sensitivity-Emulation

Creating the Conda environment

First ensure you have installed Conda. Then, check that the file "conda-environment.yml" has been downloaded when you cloned the repository. From a terminal/command line, run

conda env create -n EnvName -f conda-environment.yml 

where "EnvName" is the name of the conda environment. To match my environment, name is "flexpart". If you give it a different name, you'll need to update its name in Main.R and Train-CVAE.R.

Data files

There are a number of data files which accompany the scripts in this repository. You can access and download the data here. They are broken into three groups:

FLEXPART simulations

These are the raw output files from FLEXPART. They are broken into four groups, based on the region over which the simulations were performed: Australia, Canada, Europe, or the UK. There are 5000 simulations per region, and a 5-6 month time period over which to perform the simulations was allocated to each region. Simulations were performed at randomly drawn spatial locations and time points (within the allocated time frame). There is one output file per simulation. In each group of files, there are also three R-datasets. Two contain the post-processed, cumulated sensitivy plumes both before and after translation and rotation, and the third contains the coordinates of the spatial grid over which the simulations were performed.

Information on how to reproduce these FLEXPART outputs can be found within the FLEXPART folder in this repository.

NAME simulations

These are the cumulated sensitivity plumes from NAME, simulated over the UK and Ireland from January to April 2014. Simulations were performed every 2-hours at four measurement sites. There is one .txt file per measurement site, and per month. There are also three R-datasets. Two contain the post-processed, cumulated sensitivy plumes both before and after translation and rotation, and the third contains the coordinates of the spatial grid over which the simulations were performed.

R Datasets created during pre/post-processing and emulation

This file contains the remaining datasets not included in the NAME and FLEXPART files, produced at each stage in the pre/post-processing and emulation.

Scripts

There are a number of scripts included in this repository. Further details of their purpose can be found in "readme.md" contained in the src folder of this repository.

To reproduce results and final plots in the paper

• Before running any of the scripts, please ensure the required R packages are installed. Also, ensure you load the conda environment before installing the packages. To do this, run

install.packages("reticulate")
use_condaenv("flexpart", required = TRUE) # Change "flexpart" to the name of your own Conda environment 

• Now you can check for and install the necessary packages. The required packages for this work are reticulate, tensorflow, keras, dplyr, tidyr, gstat, fields, RNetCDF, ggplot2, gridExtra, and if running a Linux operating system, parallel. To load an already installed package (or check if a package is already installed), type

library(<package_name>)

into the console and hit the enter/return button. If the package is not yet installed, the Console will return a message saying that the package could not be found. To install a package, type

install.packages("<package_name>")

into the Console, and hit the enter/return key.

• The scripts included in this repository encompass the entire process from processing FLEXPART and NAME outputs, right through to the calculation of metrics after emulation. To recreate this process, a number of scripts need to be run in succession: get-FLEXPART-plumes-for-training.R, get-NAME-plumes-for-training.R, create-rotated-plumes-for-training.R, create-rotated-plumes-for-application.R, create-trianing-test-validation.R, create-CVAE-basis.R, create-EOF-basis.R, train-CVAE.R, train-EOFs.R, Main.R. To skip through to reproducing the plots and metrics from the paper, you need only run Main.R.

• Please note, however, that if you choose to train the CVAE yourself, the end results will be different. If you want to re-create our results exactly, please do not re-run "train-CVAE.R".

• To run each script, complete the following:

• Open a terminal/command line, and navigate the working directory to the "Data" folder (ensure you have downloaded and extracted the R Datasets within this folder).

• Activate the conda environment by running

conda activate EnvName

where EnvName is the name you gave your conda environment when creating it.

• Now open the script you want to run, and, if needed, adjust the name of the conda environment at the top of the script to match the name of your conda environment. You may also need to adjust the number of cores to suit the machine you are using. Save and close the file.

• In the activated conda environment, run, for example,

Rscript ../src/Main.R

to perform the emulation, Monte Carlo sampling and rotation, calculation of metrics, and production of plots. The final metrics should match those given in the paper (provided you have not retrained the CVAE), and your final plots should look like these:

Software license

Reproducible code for "Emulation of greenhouse-gas sensitivities using variational autoencoders", by Laura Cartwright, Andrew Zammit-Mangion, and Nicholas M. Deutscher.
Copyright (c) 2021 Laura Cartwright
Author: Laura Cartwright (lcartwri@uow.edu.au)

This program is free software; you can redistribute it and/or modify it under the terms of the GNU General Public License as published by the Free Software Foundation; either version 2 of the License, or (at your option) any later version.

This program is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for more details.