Developing Intelligent Earth System Models : An AI scheme of K-profile parameterization and stable coupling into CESM with FTA

Abstract

Parameterization schemes in numerical models are employed to represent the effects of subgrid-scale physical processes but are often limited by incomplete understanding of physical processes and computational constraints, leading to inaccuracies and inefficiencies. Artificial intelligence (AI) models have been introduced to enhance simulation accuracy or computational efficiency. However, hybrid Earth System Models (ESMs), which integrate AI into traditional frameworks, must also consider stability in coupled simulations. In this study, we replace the default K-profile parameterization (KPP) in Community Earth System Model (CESM) with a transformer-based AI model (KPP-DL). We first perform offline evaluations, demonstrating the AI model’s ability to closely replicate KPP’s key outputs. Subsequently, we couple KPP-DL into CESM via Fortran-Torch adaptor (FTA) and evaluate the hybrid CESM’s performance in terms of accuracy, stability, and computational efficiency. Hybrid CESM maintains stable operation for at least 3 years, with approximately a 3-5 times improvement in the computational efficiency of vertical mixing. During online coupled simulations, KPP-DL exhibits strong agreement with KPP in simulating key vertical mixing coefficients while hybrid CESM produces consistent results for variables such as temperature and salinity. Our results highlight the potential of AI-driven approaches to achieve accuracy approaching that of KPP and stability coupled in ESMs while improving the efficiency, suggesting that intelligent ESMs represent a promising future direction for numerical modeling.