Modelling Future Evolution of the Antarctic Ice Sheet With Explainable Machine Learning

Area of Science:

• Climate Science
• Glaciology
• Machine Learning

Background:

• Climate change accelerates ice sheet melt, leading to rising sea levels.
• Traditional ice sheet models (e.g., PISM) are computationally intensive and make simplifying assumptions, limiting long-term predictions.
• Accurate ice sheet dynamics prediction is crucial for understanding climate change impacts.

Purpose of the Study:

• To develop machine learning (ML) models that emulate ice sheet dynamics for faster and more efficient predictions.
• To improve the accuracy of ice sheet property predictions (ice mask, thickness, velocity) using ML.
• To enhance the explainability and trustworthiness of ML models in climate science.

Main Methods:

• Utilized data generated by the Parallel Ice Sheet Model (PISM) for the Antarctic ice sheet (2015-2100).
• Developed proof-of-concept ML models, including a Random Forest with an expanding window, to predict ice mask, thickness, and velocity.
• Incorporated spatial (neighboring cells) and temporal (past values) data to improve model accuracy.
• Applied Explainable AI (XAI) techniques for model interpretability.

Main Results:

• The best-performing Random Forest model achieved high accuracy: RMSE of 3.36 for thickness and 65.6 for velocity.
• The model misclassified only an average of 3 grid points per year for the ice mask.
• Achieved comparable accuracy to PISM at a significantly reduced computational cost.

Conclusions:

• ML emulation offers a computationally efficient alternative to traditional ice sheet models.
• This approach enables more frequent and extensive analyses of climate change impacts on Antarctic ice sheets.
• The study demonstrates the potential of ML and XAI in advancing climate science and prediction capabilities.