A Multimodal Earth Digital Twin Framework with Causal AI for Autonomous Climate Intervention Planning

Abstract

Climate change is among the most pressing challenges of the twenty-first century, demanding decision-support tools that are scientifically grounded, data-rich, and capable of evaluating the consequences of interventions rather than merely forecasting trends. Existing AI-driven climate platforms largely rely on correlation-based prediction models that cannot distinguish causal drivers from statistical associations, limiting their usefulness for policy evaluation. This paper proposes the Multimodal Earth Digital Twin Framework with Causal AI (MEDT-CAI), which integrates satellite imagery, climate simulation outputs, IoT sensor streams, and socio-economic indicators into a continuously updated digital twin, and augments it with a causal AI layer built on structural causal models, PCMCI+ causal discovery, and docalculus counterfactual reasoning. A multi-objective reinforcement learning engine then recommends Pareto-optimal climate interventions under environmental, economic, equity, and sustainability constraints. Evaluated on three regional case studies (Amazon Basin, Indo-Gangetic Plain, European Alps), MEDT-CAI reduces prediction error by up to 23.4% over physicsinformed baselines, improves causal discovery structural Hamming distance from 7.8 to 3.2, and achieves 84.3% expert agreement on recommended interventions, versus $\mathbf{6 1. 2} \%$ for conventional multi-criteria decision analysis. These results demonstrate that embedding causal reasoning within a multimodal Earth digital twin yields more robust, interpretable, and policyrelevant decision support than predictive-only alternatives.