Will Artificial Intelligence supersede Earth System and Climate Models?

Abstract

We outline a perspective of an entirely new research branch in Earth and climate sciences, where deep neural networks and Earth system models are dismantled as individual methodological approaches and reassembled as learning, self-validating, and interpretable Earth system model-network hybrids. Following this path, we coin the term ”Neural Earth System Modelling” (NESYM) and highlight the necessity of a transdisciplinary discussion platform, bringing together Earth and climate scientists, big data analysts, and AI experts. We examine the concurrent potential and pitfalls of Neural Earth System Modelling and discuss the open question whether artificial intelligence will not only infuse Earth system modelling, but ultimately render them obsolete. 1 ar X iv :2 10 1. 09 12 6v 1 [ st at .M L ] 2 2 Ja n 20 21 For decades, scientists have utilized mathematical equations to describe geophysical and climate processes and to construct deterministic computer simulations that allow for the analysis of such processes. Until recently, process-based models had been considered irreplaceable tools that helped to understand the complex interactions in the coupled Earth system and that provided the only tools to predict the Earth system’s response to anthropogenic climate change. The provocative thought that Earth system models (ESMs) might lose their fundamental importance in the advent of novel artificial intelligence (AI) tools has sparked both a gold-rush feeling and contempt in the scientific communities. On the one hand, deep neural networks have been developed that complement and have started to outperform the skill of process-based models in various applications, ranging from numerical weather prediction to climate research. On the other hand, most neural networks are trained for isolated applications and lack true process knowledge. Regardless, the daily increasing data streams from Earth system observation (ESO), increasing computational resources, and the availability and accessibility of powerful AI tools, particularly in machine learning (ML), have led to numerous innovative frontier applications in Earth and climate sciences. Based on the current state, recent achievements, and recognised limitations of both process-based modelling and AI in Earth and climate research, we draw a perspective on an imminent and profound methodological transformation, hereafter named Neural Earth System Modelling (NESYM). To solve the emerging challenges, we highlight the necessity of new transdisciplinary collaborations between the involved communities. Overview on Earth System Modelling and Earth System Observations Earth system models (ESMs) combine process-based models of the different sub-systems of the Earth system into an integrated numerical model that yields for a given state of the coupled system at time t the tendencies associated with that state, i.e., a prediction of the system state for time t + 1. The individual model components, or modules, describe sub-systems including the atmosphere, the oceans, the carbon and other biogeochemical cycles, radiation processes, as well as land surface and vegetation processes and marine ecosystems. These modules are then combined by a dynamic coupler to obtain a consistent state of the full system for each time step. For some parts of the Earth system, the primitive physical equations of motion are known explicitly, such as the Navier-Stokes equations that describe the fluid dynamics of the atmosphere and oceans (Fig. 1). In practise, it is impossible to numerically resolve all relevant scales of the dynamics and approximations have to be made. For example, the fluid dynamical equations for the atmosphere and oceans are integrated on discrete spatial grids, and all processes that operate below the grid resolution have to be parameterised to assure a closed description of the system. Since the multi-scale nature of the dynamics of geophysical fluids implies that the subgrid-scale processes interact with the larger scales that are resolved by the model, (stochastic) parameterization of subgrid-scale processes is a highly non-trivial, yet unavoidable, part of climate modelling. For other parts of the Earth system, primitive equations of motion, such as the Navier-Stokes equations