Source-linked AI summary
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
TL;DR
The paper addresses persistent limitations in Earth system models and the contrasting lack of physical process knowledge in purely data-driven ML. It proposes combining both approaches into learning, self-validating, and interpretable hybrids, termed Neural Earth System Modelling, while examining their potential and pitfalls. The authors conclude that hybrids can improve prediction of out-of-distribution samples and extremes and produce physically constrained, consistent simulations, whereas AI alone remains unthinkable in the current stage.
Problem
Earth system models face unresolved parameterization, abrupt-transition, and extreme-event challenges, while current AI approaches lack physical process knowledge.
Method
The paper proposes fusing process-based Earth system models with deep neural networks into learning, self-validating, and interpretable model-network hybrids.
Results
Hybrids can better reproduce and predict out-of-distribution samples and extreme events while performing constrained and consistent simulations that obey physical conservation laws.
Takeaways & Limitations
Neural Earth System Modelling is proposed as a distinct research branch for combining process-based modelling and AI in Earth and climate science.
Takeaways & Limitations
AI alone remains unthinkable in the current stage, and stochastic subgrid-scale parameterization remains unavoidable in climate modelling.
Abstract
from arXiv · showhide
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.
Overview on Earth System Modelling and Earth System Obser-
Earth system models integrate coupled Earth-system components but must approximate unresolved and poorly understood processes through parameterizations. Despite major advances, persistent uncertainties constrain predictions of climate sensitivity, abrupt changes, mitigation impacts, and extremes.
- Earth system models combine process-based modules for subsystems such as the atmosphere, oceans, biogeochemical cycles, land surface, and vegetation through dynamic coupling.
- Unresolved subgrid processes interact with larger resolved scales, making stochastic parameterization a difficult but unavoidable part of climate modelling.
- Parameterizations introduce free parameters that are often tuned manually because the size of state-of-the-art models limits systematic calibration methods.
- The likely equilibrium climate sensitivity range widened from 2.1–4.7°C in CMIP5 to 1.8–5.6°C in CMIP6, leaving future projection uncertainty unresolved.
- Current ESMs remain limited in predicting abrupt climate transitions, evaluating carbon dioxide removal, representing key environmental interactions, and reproducing extreme events.
From Machine Learning-based Data Exploration Towards Learn-
Machine learning has progressed from exploratory analysis of Earth system observations toward prediction, emulation, and parameterization of geophysical processes. The paper presents hybrids combining process-based models and ML as a foundation for Neural Earth System Modelling.
- ML has been applied to Earth observations for pattern recognition, clustering, remote-sensing analysis, time-series prediction, and climate-information reconstruction.
- Purely data-driven weather-prediction networks are beginning to compete with process-based model forecasts across spatial and temporal scales.
- ML can map nonlinear processes, but trained neural networks lack physical process knowledge because they identify statistical relations by minimizing task-specific loss measures.
- Combining ML with process-based modelling distinguishes physics-aware applications from physics-blind observation analysis and supports learning aspects of Earth and climate physics.
- ML has been used with ESMs and observations to estimate ocean heat content, recover terrestrial water storage, and upscale carbon-flux measurements.
- Hybrid and ML-based parameterization approaches can represent subgrid processes and help reduce numerical and human-induced simplifications and biases in ESMs.
- The proposed fusion defines Neural Earth System Modelling as a distinct research branch involving physics-aware ML and model-network hybrids.
The Fusion of Process-based Models and Artificial Intelligence
The paper frames Neural Earth System Modelling as a distinct research branch that fuses process-based models, AI, and observations into dynamically interacting hybrids. It emphasizes both the potential for improved prediction and the need to address stability, distribution shift, interpretability, and interface challenges.
- Weak hybrids: Neural networks can emulate unresolved or sub-grid processes using high-resolution simulations or relevant observations, but their integration may destabilize Earth system models.The choice of AI technique can significantly deteriorate numerical stability, so learned parameterizations require stabilization and evaluation within the model-physics context.
- Weak hybrids: AI components can also learn from model states, trajectories, seasonal signals, coupling mechanisms, and geophysical processes, including extreme-event prediction and possible physical causation.Earth system observations can provide additional training constraints and independent self-evaluation measures.
- Methodological caveats: Training-distribution success does not resolve the challenge of out-of-distribution prediction, making purely data-driven AI unsuitable for accurate climate projections on its own.The paper links this limitation to the non-stationarity of climate and Earth-system distributions and argues for strongly coupled hybrids and less-constrained training approaches.
- Hybrid methodological approaches: Neural Earth System Modelling combines process-based models, neural networks, and Earth system observations into increasingly strongly coupled hybrids.Weak hybrids exchange information in one direction, whereas strong hybrids dynamically exchange information through fully coupled model-network combinations.
- Goals and outlook: Future NESYM hybrids are intended to reproduce extremes, obey conservation laws, self-validate and self-correct, and improve replicability and interpretability.Achieving these goals requires innovative interfaces for information exchange that are not yet available.
Peering into the Black Box
The paper presents explainability and interpretability as central requirements for applying machine learning to climate science. XAI can expose learned relationships and evaluate climate-model behavior, but these methods remain immature and require close collaboration with climate scientists.
- Interpretability by design: Interpretable AI builds interpretability into machine-learning models from the outset, rather than explaining predictions only through post-processing diagnostics.Unsupervised approaches can expose statistically dominant physical balances and discover different system regimes.
- Why interpretability matters: Explainability makes machine-learning outputs more transparent and helps establish trust, which is important for policy-relevant climate applications.The paper stresses that researchers need to understand whether models obtain correct answers for physically meaningful reasons.
- Why interpretability matters: XAI is needed for predictions under continually evolving climate distributions because it can help explain model skill and support trust in extrapolation to future regimes.It can also assist in selecting architectures, inputs, and outputs while allowing scientists to incorporate physical knowledge into the method.
- Scientific discovery: Layerwise relevance propagation has revealed climate variability modes, predictability sources across timescales, and climate-change indicator patterns.These applications show how machine learning can support scientific questions beyond prediction alone.
- Model evaluation: XAI methods can evaluate climate models against observations and identify the most important model biases for a specific prediction task.The paper notes that these methods are still in their infancy and have substantial room for advancement.
- Transdisciplinary development: NESYM development requires climate and AI scientists to create geophysical-consistency benchmarks and address adversarial examples, artifacts, and shortcut learning.The paper concludes that continued joint development of Earth system modelling and AI is necessary for the field to evolve.
Authors’ contributions
The authors describe a collaborative perspective assembled across multiple Earth-system, observation, machine-learning, and black-box interpretability sections.
- Authors’ contributions: CI conceived the paper and organized the collaboration, while all authors contributed to writing and revision.Contributors drafted sections covering Earth system models, observations, learning physics, process-model and AI fusion, and black-box analysis.