Correlation to Causation: A Causal Deep Learning Framework for Arctic Sea Ice Prediction

TL;DR

The paper addresses the challenge that correlation-based ML/DL struggles to infer causality in Arctic sea ice forecasting. It proposes a causality-driven framework that uses Multivariate Granger Causality (MVGC) and PCMCI+ to identify causal features from 43 years of ocean‑atmospheric data and integrates them into a hybrid GRU-LSTM model. Empirical results show improved predictive accuracy and interpretability across lead times of up to six months, with different causal-feature strategies offering strengths at different horizons. The approach provides a scalable, generalizable methodology for causal deep learning in high‑dimensional dynamic systems and has potential applications beyond Arctic SIE forecasting.

Abstract

Traditional machine learning and deep learning techniques rely on correlation-based learning, often failing to distinguish spurious associations from true causal relationships, which limits robustness, interpretability, and generalizability. To address these challenges, we propose a causality-driven deep learning framework that integrates Multivariate Granger Causality (MVGC) and PCMCI+ causal discovery algorithms with a hybrid deep learning architecture. Using 43 years (1979-2021) of daily and monthly Arctic Sea Ice Extent (SIE) and ocean-atmospheric datasets, our approach identifies causally significant factors, prioritizes features with direct influence, reduces feature overhead, and improves computational efficiency. Experiments demonstrate that integrating causal features enhances the deep learning model's predictive accuracy and interpretability across multiple lead times. Beyond SIE prediction, the proposed framework offers a scalable solution for dynamic, high-dimensional systems, advancing both theoretical understanding and practical applications in predictive modeling.

Figures (6)

• Figure 1: Approximately 38% decline in Arctic September sea ice, from 2.7 million square miles in 1979 to 1.7 million square miles in 2023 (Source: US Global Change Research Program).
• Figure 2: Illustration of how causal discovery algorithms identify lagged causal relationships from time series data. $\tau$ represents the timelag of the causal links.
• Figure 3: Overview of the proposed causal deep learning framework for Arctic Sea Ice Extent (SIE) prediction. Daily and monthly temporal ocean-atmospheric datasets are analyzed with MVGC and PCMCI+ to identify causal features. These features are integrated into a GRU-LSTM architecture for SIE forecasting and performance is evaluated on causal vs. correlated features.
• Figure 4: Causal graphs generated by PCMCI+ for (a) daily and (b) monthly datasets, illustrating the causal relationships between ocean-atmospheric variables and SIE.
• Figure 5: $R^2$ values for daily models across the lead times.