Sustainable AI framework for atmospheric CO₂ forecasting using carbon-optimized self-organizing networks

Climate systems are inherently dynamic, so predicting atmospheric CO₂ for long time spans is a challenging problem due to nonlinear interactions in the environment and non-stationary behaviour of the atmosphere. We present in this work a computationally efficient and interpretable framework that accurately predicts atmospheric CO₂. This model combines the Carbon-Aware Entropic Migration Optimizer (CAEMO) which is a resource-aware convergent and stable feature refinement method with the carbon-trend based Self-Organizing Lattice Network (CaSOL-Net) which is designed for adaptive regime-based forecasting. For evaluation, we employed high-resolution atmospheric microclimate data that was gathered during the years 2018–2025, which included over 400,000 samples of temperature, humidity, pressure, radiation, wind, and atmospheric CO₂ concentration. CAEMO balanced prediction accuracy, information relevance, and computational cost to select small and informative feature subsets while CaSOL-Net captured environmental structure by applying local forecasting regimes. The results show that the proposed solution can produce better forecasting stability, eliminate redundant features, improve interpretability, and generate more meaningful regime-level CO₂ predictions compared to the classical optimization methods, which enable support of transparent climate monitoring and sustainable environmental policymaking.