This project, submitted as my honors-year research project, endeavoured to utilize evolutionary machine learning algorithms to more efficiently predict changes in Antarctic ice sheets.
Warning
This project is no longer active, and listed as an archive.
Climate change is causing sea levels to rise significantly, which considerably influences the human economy and species' living. One of the direct mechanisms that contribute to the observed sea level rise is the melting of land ice. It is vital to have some insight into how these ice sheets changes in the future. Numerical models have been used to predict how fast ice sheet is melting. These numerical models are based on complex physical numerical models with numerous complex differential equations, which is computationally expensive and time-consuming. This project will develop new evolutionary learning models that can predict ice sheet change effectively and efficiently.
After cloning the repository or downloading a zip version, the project requires several libraries to be installed (which are detailed in requirements.txt), this can be done with the following command:
After cloning the repository or downloading a zip version, the project requires several libraries to be installed (which are detailed in requirements.txt), this can be done with the following command:
pip install -r requirements.txtThis project uses datasets provided by Professor Nicholas Golledge.
The datasets are results from current physical simulations, split into .txt files labeled by year (ranging 86 years from 2015 to 2100). Each file represents 1 year, holding 8 features split across 3 types:
-
POSITIONAL CONSTANTS
Positional data is encoded with two input features:
x coordinateandy coordinate. These are constant over time, and unique to each sample. -
BOUNDARY CONDITIONS / INPUT FORCINGS
There are three temporally-evolving boundary conditions that can be used for model predictions:
precipitation,air temperature, andocean temperature. -
OUTPUTS
There are three outputs to be predicted:
ice thickness,ice velocity, andice mask. Ideally these will be predicted from the input forcings and positional data.
Several notes of domain knowledge were left with the data, giving possible expectations with how the data and models should behave according to current scientific understandings.
- Outputs are predicted to correlate most with
ocean temperaturecompared to other input features. - It is predicted that the output responses will be lagged with respect to the input forcings (ie the
ice thicknessmight start changing years or even decades after a change in boundary conditions.).
Warning
Many features hold no measured value in some samples, with these being filled with NaN or 9.96920996839e+36.
/.github- Directory containing repository management tools, including issue templates and formatting checks./data- Input datasets.EDA.ipynb- Exploratory-data-analysis (EDA)/model- Model training and testing./scripts- Helper scripts for various utilities.LICENSE- Standard MIT usage license.requirements.txt- Required Python libraries and versions for correct usage./out- Project artifacts (images, gifs, charts etc)/reports- Project proposal, preliminary report, and final report.model-testing.ipynb- Initial model testing.
All project development completed by Sam Mata, with supervision from Dr Bach Nguyen and Dr Bing Xue.
Datasets provided by Professor Nicholas Golledge with initial exploratory data analysis being completed by Serafina Slevin.
This project was completed under the Center for Data Science and Artificial Intelligence in collaboration with the Antarctic Research Centre at Victoria University of Wellington.