Interpretable foundation models as decryptors peering into the Earth system

The processes of the Earth system drive interactions between energy, matter, and life, and a comprehensive understanding of their full evolutionary trajectory is critical for sustainable human development. Traditional modeling primarily relies on a set of theoretical equations to simulate dynamic process such as the carbon-nitrogen cycle, solar radiation dynamics, and terrestrial ecosystem dynamics.1 Despite the extensive modeling experience of Earth scientists, the rapid advancement of Earth observation techniques has led to a significant increase in the volume of databases, with data accumulating daily or even hourly. This has exacerbated the conflict between the capacity for data collection and utilization for big Earth data. Consequently, there is an urgent need to enhance the intelligent processing and analysis of big Earth data.2

At this critical juncture, the emergence of foundational models3 has revitalized the unique advantages of maximizing information retrieval and deriving insights from big Earth data. However, the mathematical principles underpinning their success are somewhat elusive, raising concerns about trustworthiness due to the lack of a clearly defined internal chain of reasoning and decision-making processes. Therefore, interpretable foundational models are crucial. They enhance our understanding and security of geoscientific applications, break through performance limitations, and improve the controllability of their social impacts.