Understanding a complex earth system

Our planet, and the way it operates through various Earth System processes, is changing rapidly. Can AI help us understand and project these changes in ways that allow us to act in time?

Introduction

Earth is a complex and evolving system. It’s shaped by the intricate and dynamic interdependencies between its subsystems, such as the atmosphere, biosphere, hydrosphere, and cryosphere, which are all strongly affected by human societies and their interventions.¹ The question posed here is whether artificial intelligence (AI) can play a transformative role in advancing our understanding of the Earth System.

Traditional approaches often treat different environmental and social components in separation. In contrast, AI enables the integration and analysis of a diversity of empirical data streams— ranging from integrating satellite observations and in-situ environmental measurements, to citizen science data, socioeconomic indicators, governance records, and disaster data.

Models trained on these heterogeneous and multimodal datasets can help us explore how human activities such as pollution, land-use change, resource extraction, infrastructure development, or local to global policy decisions influence key Earth System processes like the cycling of carbon, water, nutrients, and biodiversity dynamics. These processes have implications for surface-climate feedback mechanisms, and vice versa. Below, we present some potential AI contributions to address the challenges of Earth System modeling.

AI for Earth System Data

AI offers opportunities to contribute to complex Earth Systems research in innovative ways:

AI advances remote sensing analyses of the Earth System. Over the last decades, satellitebased observations have revolutionized how researchers collect data monitoring Earth and its climate. Especially when combined with in-situ data, for example via machine learning (ML) or model-data integration, novel downstream data products emerge that enable a truly multivariate analysis of global Earth System dynamics.² Data over a few decades covers a relatively short time span, especially compared to some slowmoving key Earth System processes, such as soil, ocean, or ice sheet dynamics. Nevertheless, the data offers ample opportunities to increase our understanding of the Earth System. Training AI models on such high-dimensional data can play a crucial part in monitoring changes in data streams in time, including forecasting the complex temporal dynamics of individual Earth System components such as the atmosphere, ocean, and land,³ or by augmenting spatial resolutions to so-called super resolutions,⁴ and thus overcoming scale limitations and effectively increasing the resolution of data and models. Cloud formations are a classic example of phenomena that cannot be predicted by current numeric models at the global scale. The formation of clouds is a localized process, whereas global numeric models have a resolution that is too coarse. It creates extreme computational challenges when attempting to run a global model at such a fine resolution where cloud formations can be predicted.⁵ Today, a critical frontier of science is using AI to understand and anticipate the impacts of climate extremes.⁶ However, despite all challenges we still face, the increased data collection capabilities contribute to the advancement of predicting spatiotemporal Earth System dynamics, as in the examples provided earlier, and using them as references for process-based models.

AI provides the capacity to detect new patterns. One example where AI proved helpful is the transfer from computer vision to analyzing satellite images. These approaches lead to new and efficient solutions for classical remote sensing applications, such as monitoring land cover and land-use change,⁷ and analyzing and classifying vegetation types⁸ or land surface deformations, such as landslides.⁹ In-painting techniques from computer vision (e.g., filling in missing or corrupted parts of an image) can be used to reconstruct missing climate data and improve our understanding of the climate of the past.^10,11

AI contributes to understanding interactions and non-linear relationships. Today, big hopes are placed on interpretable and explainable AI methods (IAI and XAI, respectively) that inspect and analyze trained AI models. The idea is that if we understand what leads an AI model to predict a certain outcome, it would likewise increase our understanding of the often complex and non-linear relationships of the Earth System. Typical applications are identifying dominant driver variables or precursors of events and mechanisms that were previously not considered.^12–14 However, most methods deployed for XAI do not control for confounding factors and can be far from true causal inference. By today, the standard XAI methods remain correlative by nature but have still contributed to improved understanding of several Earth System processes. For example, XAI can help to identify the drivers and impacts of climate change on, for example, carbon dynamics,¹⁵ vegetation,¹⁶ urban temperature distributions,¹⁷ dynamic regimes in ocean modeling,¹⁸ and other domains.

AI and Predictive Modeling of the Earth System

Earth System models (ESMs) are the primary tools used to investigate the planet’s climate and how it is evolving. Predictive ESMs are numerical models that are composed of several different components, each modeling the individual parts of the Earth System (e.g., atmosphere, ocean, land) and their interactions. Traditional ESMs rely on physics-based equations and parameterizations. The parameterizations model processes that are not directly resolved by the physics-based equations, for example because they occur at smaller scales than the resolution of the model represents. However, these equations and parameterizations are computationally very expensive, and need to be improved to reduce uncertainties and biases. Such improvement is particularly important for the representation of extreme events and Earth System components with the potential of reaching tipping points, such as the polar ice sheets, the ocean current system Atlantic Meridional Overturning Circulation (AMOC), or the Amazon rainforest.¹⁹ Potential improvements to ESMs include:

Enhanced capacities for modeling and analyzing complex Earth System processes. AI may help reveal previously unknown linkages, cascading effects, and early-warning signals. For instance, ML can help identify cascading interactions between Earth System tipping elements such as the AMOC and the Amazon rainforest in observational data.²⁰ In doing so, AI strengthens our capacity to anticipate tipping points, for example in the AMOC (Theme box 2, Using AI to Detect Earth System Tipping Points), assess sustainability trade-offs, and support decisions and develop strategies that align human development within planetary boundaries.²¹

Emulating Earth System models using AI. AI can be used to accelerate existing ESMs. While ESMs, and in particular the correct setting for each parameter within and between each sub-process ESMs simulate, are computationally very expensive, AI methods can be executed efficiently on state-of-the-art GPU-supported hardware once they are trained. Therefore, deep neural networks (DNNs) can be trained to be emulators of existing models or of components of existing ESMs. In this case, the DNN is trained to exactly mimic the input–output relationship of the ESM. For example, ACE2²² is an atmospheric model emulator trained on ERA5 reanalysis data and the SHieLD model that shows an almost hundredfold increase in computational speed and decrease in energy usage compared to the process-based model. On the other hand, analogously to super-resolution exercises in remote sensing, DNNs can be integrated into ESMs to emulate subgrid-scale processes, such as those related to cloud physics and precipitation.²³ Emulating these costly subgrid-scale processes can also massively accelerate ESMs at inference. These potential performance benefits of DNNs at inference do however come with the caveat of the expensive training of DNNs. A growing number of foundation models offer new possibilities to analyze a changing Earth System (see Theme box 1, Foundation Modeling in Climate and Sustainability Science).

AI for post-processing and impact assessment. AI methods are also particularly helpful for post-processing data from ESMs. Generative AI methods inspired by computer vision and image generation have been used for downscaling and bias correction of model output data (e.g., precipitation fields) that are needed for climate impact assessment.^24,25

AI in weather prediction. Purely data-driven DNN models have also made rapid advancement in weather prediction in recent years.^26–28 These models achieve similar accuracy to physics-based weather forecasts with respect to benchmarks,²⁹ while being computationally more efficient at inference. Studies have also showcased the potential of DNN models for seasonal forecasts up to three months in advance.^30,31 Impactful climate phenomena like the El Niño– Southern Oscillation can even be forecasted up to two years in advance with AI methods.^32,33

Limitations

Some key challenges in evaluating complex Earth System research with AI include:

Challenges for long-term AI climate projections.

While the AI weather prediction systems are remarkable successes, the application of purely data-driven AI models to long-term climate projections is limited by the lack of observational data. Another problem is the intrinsic lack of internal physical consistency of AI-based predictions. High-quality and high-resolution data have only been available for a few decades,³⁴ which is short compared to the time scales of many natural variability modes of the Earth System. Naturally, there are also no observations of future climate changes from which AI methods could learn. The current generation of AI-based weather models has demonstrated surprisingly strong capabilities in generalizing to previously unseen, warmer climate scenarios. However, these models are also considerably biased toward colder climate conditions in such scenarios. However, they are also considerably biased toward a colder climate in these scenarios.³⁵ Analogous models for other critical components in the Earth System such as vegetation dynamics and their feedback mechanisms have not yet been developed.

Hybrid models: merging physics and ML. Hybrid approaches that combine process-based ESMs with DNNs seem particularly promising for the use of AI in long-term climate projections. The model elements that are based on empirical relations, such as many biological feedbacks, or are computationally extremely expensive, are learned by a deep learning model. Meanwhile, the elements whose physics are already understood by science, such as the dynamical cores of fluid dynamics in atmosphere and ocean models, are inherited from the ESM.³⁶ Differentiable programming that leads to models that are both objectively calibratable on observational data and in which DNNs can be more easily integrated are a potential perspective for such models.³⁷ One such differentiable, hybrid model is NeuralGCM.³⁸ It is an atmospheric model that combines physics-based fluid dynamics with an AI parameterization. NeuralGCM is competitive to purely physics-based models for weather prediction and can be used for decadal simulations. Hybrid models have both improved stability and short-term predictability compared to purely data-driven DNN approaches.³⁹ Further work has also shown the potential of hybrid models for the global hydrological cycle^40,41; for the modeling of turbulence, convection, and radiation in the atmosphere, and of fire dynamics in land models42; and for bias correction of sea ice modeling.⁴³ While a new generation of automatic differentiation tools such as JAX and Enzyme^44,45 makes writing such hybrid models easier, they do usually need a complete rewriting of the physics-based code, which is a considerable hurdle in their adaptation.

Potential

Finally, there are many promising sustainability science areas in which AI methods can augment our understanding of the Earth System. Some examples include modeling and scenario-building of high-risk phenomena, such as climate tipping dynamics and Earth resilience loss, enhanced Earth monitoring, and incorporating the increasing amount of data that we collect into our models. Physics-based and classical statistical models will remain instrumental in Earth System modeling, especially when making climate projections into future scenarios. AI can augment those models efficiently and leverage available data.

THEME BOX 2

Using AI to Detect Earth System
Tipping Points

Authors: Juan C. Rocha, Jonathan F. Donges, Ingo Fetzer, Maximilian Gelbrecht

Tipping points are defined as critical thresholds in the parameters of a system, transgressions of which can trigger a substantial qualitative change in the behavior of its long-term dynamics, often driven by amplifying feedback mechanisms.⁴⁷ Key large-scale examples in the Earth System include the irreversible collapse of major ice sheets, the dieback of the Amazon rainforest into a savannah, and the shutdown of major ocean current systems like the Atlantic Meridional Overturning Circulation (AMOC). In the field of mathematics, these phenomena are known as bifurcations, referring to the characteristics of nonlinear dynamics.⁴⁸ Hysteresis, defined as the degree of the reversibility of these shifts, is a sufficient but not necessary condition for tipping points. This limitation is demonstrated by the existence of continuous tipping phenomena, such as pandemics, the disappearance of Arctic sea ice, and certain social diffusion processes. Artificial intelligence (AI) and machine learning (ML) methods are well suited to learn complicated nonlinear functional forms from data. Hence, these methods have been proposed as potential promising future avenues to detect, predict, and anticipate tipping points–through early warning signals–in a variety of systems relevant to sustainability and Earth System science. These include, for example, climatic, biological, social, and social-ecological systems.^49-51 The AMOC is an important part of the Earth’s climate system. Mounting scientific evidence shows that this system of ocean currents possesses potential tipping points.⁵² These potential tipping points have, in turn, been scrutinized using deep learning (DL),⁵³ reservoir computing,⁵⁴ or ML-based rare event detection techniques.⁵⁵

Initial experiments in detecting and predicting tipping points with ML have mainly used deep neural networks, support vector machines, or random decision forest approaches. The ML techniques have been applied so far mostly on synthetic data from computer simulations, but only scarcely on real-world data.^49-51,56-58 Due to the enormous learning skills of ML, tipping points in artificial data can be successfully detected. However, it should be noted that being based on computer simulations, these approaches are constrained by certain assumptions of an upcoming and anticipated bifurcation, and most significantly about their well-defined and often low dimensionality. These assumptions impose limitations on the applicability of these approaches in real-world systems as it is not known if a bifurcation actually exists and knowledge on their dimensionality is hypothetical.⁵⁹

Recent efforts have attempted to incorporate novel approaches such as deep neural networks for higher-dimensional bifurcations.⁶⁰ Efforts have also been made to include a more diverse array of pathways, more likely leading toward bifurcation, such as rate-induced or noise-induced tipping.⁶¹ Even more recent efforts have focused on the prediction of real observed abrupt transitions from observational data with promising outcomes for future research.⁶² One key challenge in implementing and scaling up these ML methods for the prediction of tipping points is the scarcity of observed annotated datasets from real systems on which the models can be trained and evaluated. Consequently, until such datasets become available, the evaluation of precision and reduction of uncertainty will remain very much constrained.