Developing intelligent Earth System Models: An AI framework for replacing sub-modules based on incremental learning and its application

Abstract

In recent years, developing intelligent Earth System Models (ESMs), which integrate artificial intelligence (AI) models to replace ESMs' sub-modules, has shown promising results in enhancing both the accuracy and efficiency of traditional ESMs. A crucial determinant in this process is the quality of training data, which determines what knowledge AI can learn. Delving into specifics, while training with sub-module simulation data bolsters computational swiftness, it can come at the cost of accuracy. On the other hand, training with finer-resolution data (e.g., high-resolution model simulation and observational data) augments accuracy but introduces challenges when coupling AI into ESMs, such as instability or even crashing, and deteriorating accuracy over time. Notably, sub-module simulation and finer-resolution data can complement each other and provide different knowledge to AI. Therefore, our study proposes a novel AI-RepSub framework for replacing sub-modules, which employs the incremental learning (IL) method to learn from multiple data types and preserve previously acquired knowledge when fitting new data. To verify the AI-RepSub's performance, this study chooses to substitute the convection trigger function (CTF) in the Community Earth System Model (CESM), developing a hybrid CESM with an intelligent CTF module (CESM-iCTF). Our online evaluations indicate that CESM-iCTF outperforms CESM in CTF accuracy, computational efficiency, and North Atlantic Oscillation (NAO) hindcast skill while maintaining stability. To the best of our knowledge, this represents the first exploration into AI sub-modules driven by multi-source data in the field of intelligent ESMs, thereby demonstrating the potential of AI to serve as robust replacements for ESMs' sub-modules.