Reduced Cloud Cover Errors in a Hybrid AI-Climate Model Through Equation Discovery And Automatic Tuning (2505.04358v1)

Abstract: Climate models rely on parameterizations that account for the effects of small-scale processes on large-scale dynamics. Particularly cloud-related parameterizations remain a major source of uncertainty in climate projections. While hybrid Earth system models (ESMs) with machine learning-based parameterizations could improve current ESMs, deep learning approaches often lack interpretability, physical consistency, and computational efficiency. Furthermore, most data-driven parameterizations are trained in a stand-alone fashion and fail within ESMs, partly due to the difficulty of tuning the ESM to accommodate new, non-traditional schemes. In this study, we introduce a novel two-step pipeline for improving a climate model with data-driven parameterizations. First, we incorporate a physically consistent, data-driven cloud cover parameterization into the ICON global atmospheric model. The parameterization, a diagnostic equation derived from storm-resolving simulations via symbolic regression, retains the interpretability and efficiency of traditional parameterizations while improving fidelity. Second, we introduce an automated, gradient-free tuning procedure to recalibrate the new climate model with Earth observations. We employ the Nelder-Mead algorithm and progressively increase simulation length, making our approach simple, computationally efficient, and easily extendable to other ESMs. The tuned hybrid model significantly reduces some long-standing biases in cloud cover and radiative budgets, particularly over regions such as the Southern Ocean and the subtropical stratocumulus regions. Moreover, it remains robust under +4K surface warming. Our results highlight the potential of data-driven parameterizations when combined with model tuning. This framework offers an automatic, efficient and practical approach to enhancing climate projections without losing performance or interpretability.