Developing explainable AI models for predicting snowpack variability in Polar regions

Accurate prediction of variability of polar snowpack is central to understanding global climate change and its impacts on ecosystems, sea-level rise, and weather patterns. Traditional, physically based snowpack models commonly fail to capture full range of complexity that realistic snow dynamics can display, especially under extreme climate events. On the contrary, machine learning (ML) models can enhance predictive accuracy but lack interpretability and thus are challenging to apply in scientific contexts. This research proposes construction of explainable AI models for prediction of polar snow-pack, emphasizing how to interpret domain knowledge coupled with state-of-the-art techniques in AI. Physics-informed neural networks are used herein to embed physical laws, Graph Neural Networks (GNNs) to represent spatial dependencies, and Recurent Neural Network (RNNs) for temporal sequences. Several post-hoc explanation techniques, such as SHAP and saliency maps, together with the development of causal inference models, are applied in such a way that the transparency and scientific basis of the models are preserved.