Understanding the Influence of Parameter Value Uncertainty on Climate Model Output: Developing an Interactive Web Dashboard

Scientists at the National Center for Atmospheric Research have recently carried out several experiments to better understand the uncertainties associated with future climate projections. In particular, the NCAR Climate and Global Dynamics Lab (CGDL) working group has completed a large Parameter Perturbation Experiment (PPE) utilizing the Community Land Model (CLM), testing the effects of 32 parameters over thousands of simulations over a range of 250 years. The CLM model experiment is focused on understanding uncertainty around biogeophysical parameters that influence the balance of chemical cycling and sequestration variables. The current website for displaying model results is not intuitive or informative to the broader scientific audience or the general public. The goal of this project is to develop an improved data visualization dashboard for communicating the results of the CLM PPE. The interactive dashboard would provide an interface where new or experienced users can query the experiment database to ask which environmental processes are affected by a given model parameter, or vice versa. Improving the accessibility of the data will allow professionals to use the most recent land parameter data when evaluating the impact of a policy or action on climate change.

Methods
Data Source:

University of California, Santa Barbara – Climate and Global Dynamics Lab, National Center for Atmospheric Research: Parameter Perturbation Experiment (CGD NCAR PPE-5). https://webext.cgd.ucar.edu/I2000/PPEn11_OAAT/ (Only public version of the data currently accessible. Data leveraged in this project is currently stored on the NCAR server and is not publicly available), https://www.cgd.ucar.edu/events/seminar/2023/katie-dagon-and-daniel-kennedy-132940 (Learn more about this complex data via this amazing presentation by Katie Dragon & Daniel Kennedy ^)
The Parameter Perturbation Experiment data leveraged by our project was generated utilizing the Community Land Model v5 (CLM5) predictions. https://www.earthsystemgrid.org/dataset/ucar.cgd.ccsm4.CLM_LAND_ONLY.html

Data Processing:

We were working inside of NCAR’s CASPER cluster HPC server, this enabled us direct access to the raw data files. We created a script to read in 500 LHC PPE simulations as a data set with inputs for a climate variable and time range. When reading in the cluster of simulations, there is a preprocess function that performs dimensional reduction to simplify the data set for wrangling later.

Once the data sets of interest were loaded, they were then ready for some dimensional corrections – some quirks that come with using CESM data. Our friend’s at NCAR CGDL actually provided us with the correct time-paring bug. The other functions to weigh each grid cell by land area, properly weigh each month according to their contribution to the number of days in a year, and to calculate the global average of each simulation were generated by our team to wrangle the data so it is suitable for emulation. These files were saved so they could be leveraged later using a built-in if-else statement within the `read_n_wrangle()` function.

The preprocessed data is then used in the GPR ML Emulator to make 100 predictions for a climate variable of interest and 32 individual parameters. To summarize briefly without getting too into the nitty gritty, our GPR emulator does 3 things:

Simplifies the LHC data so it can look at 1 parameter at a time and assess its relationship with a climate variable.
 Applies Fourier Amplitude Sensitivity Analysis to identify relationships between parameters and climate variables. It helps us see what the key influencers are.
 In the full chaotic LHC, it can assess the covariance of the parameter-parameter predictions simultaneously (this is the R^2 value you’ll see on your accuracy inset plot later)

Additionally, it ‘pickles’ and saves the predictions and trained gpr_model so they can be utilized for further analysis, exploration, and visualizations.

Attributes and structures defined in this notebook outlines the workflow utilized to generate the data in this repo. It pulls functions from this utils.py to execute the desired commands. Below we will look at the utils.py functions that are not explicitly defined in the notebook. – General side note: if you decide to explore that Attributes and structures defined in this notebook explaining how the data was made, you’ll notice you’ll be transported to another repo in this Organization: GaiaFuture. That’s our prototype playground! It’s a little messy because that’s where we spent the second half of this project tinkering. The official repository is https://github.com/GaiaFuture/CLM5_PPE_Emulator.