Will Artificial Intelligence supersede Earth System and Climate Models?

Christopher Irrgang, Niklas Boers, Maike Sonnewald, Elizabeth A. Barnes, Christopher Kadow, Joanna Staneva, Jan Saynisch-Wagner

Introduction

The rapid advancement of deep learning has sparked a debate in climate science: will artificial intelligence eventually replace traditional physical models? This paper introduces Neural Earth System Modelling (NESYM), a research paradigm that seeks to fuse the physical rigor of traditional Earth System Models (ESMs) with the pattern-recognition capabilities of machine learning. Rather than viewing AI as a replacement, the authors propose a "self-validating" hybrid approach to overcome persistent uncertainties in climate prediction while maintaining physical consistency.

The Challenge: Scale and Process Knowledge

For decades, ESMs have been the cornerstone of climate projection, relying on fundamental physical equations like Navier-Stokes for fluid dynamics. However, these models struggle with scale. Critical processes—such as cloud formation or oceanic convection—occur at subgrid-scales smaller than the model's resolution. To account for these, scientists use "parameterizations," which are simplified approximations that often introduce significant uncertainty and require manual tuning.

Symbolic representation of Earth system components and their coupling mechanisms across different spatio-temporal scales.

While modern AI excels at identifying patterns in massive datasets, it often lacks "process knowledge." A pure AI model might excel at short-term weather prediction but fail to simulate a 21st-century climate it has never encountered—a challenge known as the "out-of-distribution" problem.

Methodology: The NESYM Framework

The authors outline a roadmap for transitioning from isolated tools to integrated hybrids, categorized by the depth of their coupling:

  • Weakly Coupled NESYM: In this stage, one system informs the other. For instance, a neural network might be trained on high-resolution simulation data to replace a computationally expensive physical parameterization. Alternatively, ESM outputs can pre-train AI models to help them "inherit" geophysical causation.
  • Strongly Coupled NESYM: This represents the frontier of the field, involving fully integrated model-network combinations that dynamically exchange information. These hybrids would include adaptive measures for self-validation and self-correction, essentially "learning" during execution.
The successive stages of fusing physical equation-driven modelling with data-driven machine learning to reach the goal of Neural Earth System Modelling.

A cornerstone of this methodology is Physics-Informed AI, where physical laws (such as the conservation of mass and energy) are embedded directly into the neural network's architecture or loss functions to ensure the AI doesn't produce "unphysical" results.

Data Foundations

While not introducing a new dataset, the paper emphasizes three critical data pillars:

  1. Earth System Observations (ESOs): Satellite data and buoy arrays provide the "ground truth" for validation.
  2. Climate Model Intercomparison Project (CMIP): Large-scale multi-model datasets (e.g., CMIP6) are used to train AI surrogates and identify emergent constraints.
  3. High-Resolution Simulations: These serve as "training environments" for neural networks to learn subgrid-scale processes that are too computationally expensive for global climate runs.

Beyond the Black Box: Explainability and Discovery

For AI to be viable in climate policy, it must move beyond "black box" operations. The authors stress the need for Explainable (XAI) and Interpretable (IAI) AI. Understanding why a model predicts a specific outcome is essential for trusting its extrapolations into a warming world. Furthermore, AI should not just emulate existing models to save time; it should help discover new physical relationships or "causal links" within the Earth system that are currently poorly understood.

Current Progress and Proofs-of-Concept

The paper cites several successful applications of these hybrid concepts:

  • Improved Forecasting: ML-based models are beginning to outperform traditional models in specific domains like numerical weather prediction and ENSO (El Niño) forecasting.
  • Data Reconstruction: AI has been used to "fill in" missing historical climate data, improving the validation of long-term trends.
  • Advanced Parameterization: Neural networks have demonstrated the ability to approximate complex turbulent processes more accurately than traditional linear approximations.

Future Directions and Technical Hurdles

Despite the promise of NESYM, several challenges remain:

  • Numerical Stability: Integrating neural networks into physical models can cause numerical instabilities that crash simulations.
  • Non-Stationarity: Because human activity is changing the climate, AI models trained on historical data may struggle with future scenarios, necessitating "physical anchors" from ESMs.
  • The "Shortcut" Problem: AI often finds correlations that work in training but lack physical basis. Developing benchmarks for "geophysical consistency" is a critical next step.

Conclusion

The authors conclude that while AI is unlikely to render ESMs obsolete, the fusion of the two is both inevitable and necessary. NESYM represents a transdisciplinary bridge, requiring climate scientists and AI experts to work within a shared conceptual framework to build the next generation of climate science tools.

Summary updated

Abstract

In the past few years, AI approaches have been used to enhance Earth and climate modelling. This Perspective examines the opportunity to go further, and build from scratch hybrid systems that integrate AI tools and models based on physical process knowledge to make more efficient use of daily observational data streams.
BibTeX
@article{irrgang2021will,
  author = {Christopher Irrgang and Niklas Boers and Maike Sonnewald and Elizabeth A. Barnes and Christopher Kadow and Joanna Staneva and Jan Saynisch-Wagner},
  title = {{Will Artificial Intelligence supersede Earth System and Climate Models?}},
  journal = {Nat. Mach. Intell.},
  year = {2021},
  doi = {10.1038/s42256-021-00374-3},
}
Figure