Artificial intelligence for geoscience: Progress, challenges, and perspectives

Abstract

This paper explores the evolution of geoscientific inquiry, tracing the progression from traditional physics-based models to modern data-driven approaches facilitated by significant advancements in artificial intelligence (AI) and data collection techniques. Traditional models, which are grounded in physical and numerical frameworks, provide robust explanations by explicitly reconstructing underlying physical processes. However, their limitations in comprehensively capturing Earth’s complexities and uncertainties pose challenges in optimization and real-world applicability. In contrast, contemporary data-driven models, particularly those utilizing machine learning (ML) and deep learning (DL), leverage extensive geoscience data to glean insights without requiring exhaustive theoretical knowledge. ML techniques have shown promise in addressing Earth science-related questions. Nevertheless, challenges such as data scarcity, computational demands, data privacy concerns, and the “black-box” nature of AI models hinder their seamless integration into geoscience. The integration of physics-based and data-driven methodologies into hybrid models presents an alternative paradigm. These models, which incorporate domain knowledge to guide AI methodologies, demonstrate enhanced efficiency and performance with reduced training data requirements. This review provides a comprehensive overview of geoscientific research paradigms, emphasizing untapped opportunities at the intersection of advanced AI techniques and geoscience. It examines major methodologies, showcases advances in large-scale models, and discusses the challenges and prospects that will shape the future landscape of AI in geoscience. The paper outlines a dynamic field ripe with possibilities, poised to unlock new understandings of Earth’s complexities and further advance geoscience exploration.

Graphical abstract

Public summary

• •What does AI bring to geoscience? AI has been accelerating and deepening our understanding of Earth Systems in an unprecedented way, including the atmosphere, lithosphere, hydrosphere, cryosphere, biosphere, anthroposphere and the interactions between spheres.
• •What are the noteworthy challenges of AI in geoscience? As we embrace the huge potential of AI in geoscience, several challenges arise including reliability and interpretability, ethical issues, data security, and high demand and cost.
• •What is the future of AI in geoscience? The synergy between traditional principles and modern AI-driven techniques holds immense promise and will shape the trajectory of geoscience in upcoming years.

Introduction

Geoscientists tackle the most significant environmental, scientific, and societal challenges related to Earth.^1,2 Despite extensive research, several questions remain unanswered, such as the origin of Earth and life^3,4 or the snowball/faint sun paradox,⁵ among others.^6,7,8,9 Unraveling these mysteries requires modeling a complex geosystem,¹⁰ where Earth presents complicated spatial patterns shaped by diverse interacting processes, including natural sub-systems (such as the biosphere, atmosphere, and lithosphere) and various human activities.^11,12,13 For instance, predicting geohazards necessitates considering not only the inherent complexities of the geosystem but also the significant influence of activities across multiple spatial scales.¹⁴ Moreover, the geosystem is an ever-evolving network characterized by non-linear processes of high dynamical instability,¹⁵ where inherently stochastic features impose significant constraints on temporal analyses.¹⁶ In this context, while current weather forecasting can achieve relatively accurate predictions over several days, the challenge of making reliable predictions intensifies when extending the time frame to months or longer.¹⁷ Throughout history, geoscience has undergone a transition from the reliance on physics-based models to the utilization of data-driven machine learning (ML) approaches when tackling these challenges. This shift has been facilitated by remarkable advancements in artificial intelligence (AI),¹⁸ data collection techniques,^19,20,21 and computing resources.²²

Physics-based models

Geoscientific research fundamentally relies on conceptual models that describe key processes and their interactions,²³ which are subsequently tested using physical and numerical models.²⁴ Physical models simulate environmental conditions in a laboratory setting,²⁵ allowing researchers to manipulate variables in a controlled manner and investigate hypothetical scenarios within a large-scale and complicated geosystem.²⁴ While physical models are effective in certain cases (e.g., using clay models to verify the orogenic theory²⁶), they also encounter discrepancies between the controlled virtual laboratory environment and real-world situations.²⁷ Numerical models condense natural processes into mathematical representations.²⁸ These equations, designed to mirror the intricate characteristics of the real geosystem, are too complex to be solved analytically²⁹ and are generally tackled by numerical simulations, such as numerical weather prediction models.³⁰ Traditional physics-based models aim to uncover hidden mechanisms by reconstructing physical processes, and can provide robust explanations once successfully founded. However, the inherent complexity of the geosystem, coupled with our limited understanding, poses a significant challenge.³¹ Making comprehensive assumptions about related factors and their dependencies thus becomes difficult.^32,33 The sophistication and uncertainty in optimizing such models greatly hinders their practical application.³⁴

Data-driven approaches

As the availability of geoscience data continues to expand, modern geoscientific challenges are increasingly centered around managing extensive datasets, often with limited or no underlying theoretical knowledge.^17,35,36 In this context, AI demonstrates significant potential.^22,37,38 ML, as a major subfield of AI, is deeply rooted in applied statistics and constructs computational models based on inference and pattern recognition rather than physical rules.^39,40 Typical examples include Gaussian-process-based “Kriging” interpolation,^41,42,43 the utilization of support vector machines for identifying geomorphological features,⁴⁴ and so on.^45,46,47 The success of these methods has sparked broad interest among geoscientists in employing ML to address Earth science challenges, allowing them to bypass the explicit modeling of physical processes.^44,48,49 While conventional ML methods can effectively handle small-scale problems, they often encounter limitations in more complicated scenarios, particularly when dealing with large volumes of data and broader scales.⁵⁰ In this case, deep learning (DL) has brought significant advances^51,52,53 since AlexNet decisively won the ImageNet challenge in 2012.⁵⁴ Beyond applications of convolutional neural networks and Vision Transformers (ViTs),^55,56 densely/fully connected networks have proven useful in tasks such as soil mapping,⁵⁷ while recurrent neural networks, including long short-term memory (LSTM) networks, are particularly well suited for time series data and temporal problems.⁵⁸ AI models hold promise for advancing modern geoscientific research by learning hidden features directly from data without requiring comprehensive physical prior knowledge. DL, as the primary data mining tool in the big data era, propels the application of AI to geoscience. However, AI techniques still face several challenges, including the notorious data-hungry characteristics, the increased demand for computational resources, and the inherent black-box nature of AI algorithms.^59,60 Addressing these challenges is crucial to further explore the potential of AI in geoscience.

Advanced AI techniques

Solely relying on either physics-based or data-driven models proves insufficient for knowledge discovery in geoscience.⁶¹ Hybrid models or physics-guided/informed/aware ML offer a promising solution by integrating domain knowledge to refine AI models in geoscience.⁶² These models incorporate constraints derived from domain-specific insights, such as encoding differential equations from data⁶³ or imposing physical constraints on data-driven models.^64,65 This integration allows for performance comparable with pure data-driven approaches but with the advantage of requiring less training data.⁶⁶ Despite their potential to bridge interdisciplinary gaps between data-driven and physics-based models, the effective implementation of hybrid models remains an open question.^11,67 In addition, the recent success of ChatGPT has emphasized the potential of foundation models to enhance a wide range of tasks.⁶⁸ The vast expansion of data in geoscience provides a solid groundwork for the emergence of large geoscientific models.⁶⁹ These large models offer new avenues for extracting new insights from data to enrich our understanding of the Earth. Nevertheless, their development is still in the early stage.^70,71 Geo-data possesses unique characteristics, such as geo-references, various attribute features, and temporal constraints, which make it challenging to directly apply prominent language- and image-processing techniques from other fields to geoscience. How to formulate foundation models tailored to geoscience, with implications for diverse downstream tasks, remains an underexplored area. Furthermore, humanity’s quest for knowledge has increasingly extended beyond Earth into outer space.^72,73 The 21st century has seen significant advancements in space exploration.^74,75,76 For example, NASA’s Artemis campaign aims to explore the Moon for scientific research and technological advancement in 2024,⁷⁷ alongside China’s Chang’e program.^78,79 The BepiColombo mission of European Space Agency targets perplexing questions about Mercury, aiming to unravel the history of the entire Solar System.⁸⁰ With our knowledge of other planets still limited, advanced AI techniques play a crucial role in processing and analyzing the vast amounts of data collected from these missions. By deepening our comprehension of planetary processes, we cannot only enhance our understanding of these celestial bodies but also enrich Earth-based research by drawing insightful comparisons between fundamental geological mechanisms and planetary evolution.^81,82

Advanced AI techniques, particularly emerging paradigms such as physics-informed ML and large models, showcase unprecedented potential for advancing geoscience. These innovative approaches open new avenues for addressing complex challenges not only in Earth science but also in the exploration of outer space. However, current research in these promising domains remains relatively limited. This article aims to offer a comprehensive overview of the latest advancements in AI applications within geoscience. In addition, it discusses the associated challenges and identifies untapped opportunities in this field, providing guidance for future works. While several reviews have previously explored the application of AI in geoscience, offering valuable insights into the evolving landscape,^59,61,70,83 the rapid evolution of AI, up-to-date reviews to capture current trends and illuminate future research directions. Geoscience, in particular, requires special considerations for AI methodology design, given the unique characteristics of geo-data. Therefore, rather than revisiting fundamental concepts and exemplified applications of commonly used ML models,^59,70,83 our work highlights the latest achievements and prospects of AI, especially in handling big geoscience data. We will demonstrate how AI can overcome the trade-off between efficiency and accuracy, as well as make breakthroughs in other aspects, such as providing new plausible hypotheses and research directions. Furthermore, we summarize new emerging geoscientific questions and paradigms in the context of modern AI and contemporary space exploration to shed light on potential future avenues for geoscience researchers.

The rest of the paper is organized as follows. The section “geoscientific research paradigms” summarizes major geoscientific research paradigms, with a special focus on AI-related ones in section “AI-driven geoscience paradigms” and some typical application cases in section “typical cases”. The latest progress of geoscientific large models is demonstrated in section “large models in geoscience”. Then, we present some challenges and plausible future lines for contemporary AI geoscientific method design in section “challenges and outlooks in AI for geoscience”, followed by some findings in the “conclusion” section.

Geoscientific research paradigms

Diverse approaches and paradigms have been developed to deepen our understanding of the dynamic Earth system.⁸⁴ This section offers a comprehensive overview of the field, encompassing research paradigms ranging from traditional observational studies to advanced computational analyses. Four distinct yet interconnected methodologies have shaped contemporary geoscience: the observational-hypothesis-driven paradigm, the model-driven paradigm, the data-driven paradigm, and the model-data-driven paradigm,^10,18 as illustrated in Figure 1. Each of these methodologies brings unique strengths from foundational theories to advanced simulations and analyses, contributing to our comprehension of the Earth’s complex system through collaborative synergies.

Figure 1.

Illustration of four research paradigms in geoscience

Observational-hypothesis-driven paradigm

The observational-hypothesis-driven paradigm is foundational to Earth system science, playing a crucial role in understanding the complex interactions within our planet’s interconnected systems.^19,85 This approach involves systematic data collection and analysis to develop hypotheses about Earth’s processes, dynamics, and derived consequences. Rooted in empirical evidence and scientific methods, this paradigm emphasizes objective observation and rigorous hypotheses testing.⁸⁶ A seminal example of this paradigm is James Lovelock’s Gaia hypothesis,⁸⁷ which suggests the Earth’s biosphere functions as a self-regulating system, a concept that fundamentally requires extensive Earth system observations to be substantiated. Observations validate and refine hypotheses, providing insights into processes that may not be directly observable. For instance, the study of ocean circulation has greatly benefited from observations of sea surface temperatures and currents, which have been crucial in understanding the dynamics of events such as El Niño.⁸⁸ In addition, using empirical observations and hypothesis testing in frameworks such as the Community Earth System Model has also been instrumental in assessing future climate scenarios and informing policy formulation.⁸⁹

The advancement of technology has revolutionized our ability to collect data from various sources, including satellites, ground-based sensors, and remote sensing instruments.⁹⁰ Acquiring high-quality observational data has allowed scientists to refine their hypotheses and models, leading to more accurate predictions and a deeper understanding of Earth’s behavior. In climate science, the Intergovernmental Panel on Climate Change Assessment Reports exemplify this paradigm in action. Leveraging extensive observational data, these reports critically evaluate the current state of climate system, hypothesize about future climate trends, and predict potential global and societal impacts. This demonstrates the profound impact of systematic observations on both scientific and policy-oriented discourse.⁹¹

In summary, the observational-hypothesis-driven paradigm is a fundamental method that combines empirical observations and hypothesis testing to unravel the interactions within Earth’s interconnected systems. Firmly rooted in the scientific method, this paradigm remains indispensable for deciphering the intricate operations of the Earth system and guiding our responsible stewardship of the planet.

Future work within this paradigm should focus on advancing the integration and resolution of sensor networks across diverse ecosystems. By enhancing data collection methodologies, researchers can improve the accuracy of environmental models, leading to a more refined understanding of Earth system dynamics and their implications for global climate patterns. This approach will enable more precise predictions and foster a deeper scientific understanding of interconnected planetary systems.

Model-driven paradigm

There has been a long-standing focus on deciphering the interactions between natural processes and human activities on the Earth’s surface.⁹² This focus has driven the development of computational techniques and mathematical models, particularly process-based ones,⁹³ which simulate the physical, chemical, and biological processes of the Earth.⁹⁴ These models vary in complexity, ranging from simple representations of single processes to intricate integrations of multiple systems. The crux of process-based modeling lies in its challenges to transform our conceptual understanding of Earth processes into quantifiable and replicable frameworks.⁹⁵ By employing mathematical representations of natural phenomena, these models provide insights into the mechanisms driving Earth’s systems.⁹⁶ Examples of such models include the Soil and Water Assessment Tool⁹⁷ and Storm Water Management Model⁹⁸ for hydrological studies, and the Finite Volume Community Ocean Model⁹⁹ for oceanic processes. These models have significantly enhanced our understanding and predictive capabilities regarding natural phenomena. In atmospheric science, models such as the Weather Research and Forecasting model,¹⁰⁰ Community Multiscale Air Quality model,¹⁰¹ and Model of Emissions of Gases and Aerosols from Nature¹⁰² are particularly pivotal. The predictive power of process-based models is substantial, allowing scientists to explore “what-if” scenarios that inform decision-making in environmental management and policy.¹⁰³ However, the efficacy of these models depends on their calibration and validation against empirical data.¹⁰⁴ This iterative process of refinement and validation ensures the models’ accuracy and relevance, highlighting the continuous evolution of our understanding of Earth’s systems through scientific inquiry and computational innovation.^105,106

Future efforts in the model-driven paradigm should concentrate on refining model scalability and resolution, particularly by incorporating adaptive algorithms that improve the fidelity of simulations under varying climatic and environmental conditions in current and future scenarios. This will expand our capacity to predict subtle changes within Earth’s systems with greater precision.

Data-driven paradigm

The data-driven approach has revolutionized our understanding of Earth systems and human-environment interactions.¹⁰⁷ Fueled by the vast availability of data and advancements in computing and sensing technologies, this paradigm allows researchers to gain deeper insights into the complex interplay between natural processes and human activities.¹⁰⁸ In geoscience, this paradigm shift is exemplified by utilizing satellite imagery and all kinds of big geo-data.^109,110 For instance, the analysis of observational data has enabled researchers to monitor changes in land cover, deforestation rates, and urban expansion, providing crucial information for sustainable land-use planning and climate change.^111,112,113 Data-driven methods have also transformed our understanding of urban environments.^114,115 The analysis of transportation data, such as traces from global positioning systems and traffic flow data, has enabled researchers to model urban mobility patterns and reduce traffic congestion.¹¹⁶ In addition, social media data and geotagged content have provided insights into human behavior, sentiment, and urban cultural dynamics, shedding light on the social aspects of urban life.¹¹⁷ The data-driven paradigm has also facilitated the study of the human-environment nexus in urban areas. By integrating data on air quality, land use, and human activity, researchers can better comprehend how urbanization affects air pollution, public health, and carbon neutrality.^118,119 This holistic approach has been instrumental in shaping policies aimed at improving urban air quality and reducing pollution-related health risks. Moreover, the integration of socioeconomic and environmental data has enhanced our understanding of urban resilience and vulnerability to natural disasters.^120,121 For instance, by analyzing demographic data and flood risk maps, researchers can identify vulnerable populations in flood-prone areas and devise targeted disaster preparedness strategies.¹²²

In summary, the data-driven approach has propelled our understanding of Earth systems and urban dynamics to new heights. By harnessing vast datasets and sophisticated computational techniques, researchers can now explore the intricate connections between natural processes and human activities, facilitating more informed decision-making in areas such as land use, climate adaptation, transportation planning, and disaster resilience. This paradigm shift advances our scientific knowledge and offers practical solutions to the challenges facing our planet and urbanized societies.

Future work in the data-driven paradigm should emphasize the development of real-time data processing and analytics frameworks. By enabling instantaneous analysis and application of Earth system data, researchers can deliver more timely responses to environmental changes and disasters, thereby enhancing decision-making processes in critical situations.

Model-data-driven paradigm

The integration of process-based and data-driven models, commonly referred to as hybrid models, leverages the strengths of both paradigms and advances our comprehension of Earth system dynamics.¹⁷ Hybrid modeling enhances simulation precision and computational efficiency.¹²³ Process-based models, underpinned by equations of 171 motion, are particularly effective in capturing the processes of atmospheric and oceanic dynamics. However, they often struggle with complex areas such as biological processes and carbon cycling, where numerical methods fall short and semi-empirical methods lack the necessary details and accuracy.¹²⁴ Hybrid models address this gap by employing ML to replace empirical sub-models, utilizing extensive observational data while maintaining process-based models for well-understood mechanisms.¹²⁵ In addition, certain components of Earth system models are computationally expensive, particularly when handling large datasets involving complex partial differential equations¹²⁶ or high-dimensional problems.¹²⁷ Despite the fact that ML emulators may incur high initial training costs, they offer a significant reduction in computation time once operational, outperforming traditional local process modules.¹²⁸ This increase in computational efficiency not only accelerates model processing but also enhances sensitivity and uncertainty analyses. The data-driven aspects of these hybrid models afford the flexibility needed to adapt to evolving conditions, as seen in climate and vegetation dynamic modeling.¹⁷ Moreover, integrating physical principles into ML models enhances interpretability and extends their ability to extrapolate beyond observed datasets. For instance, domain-specific knowledge and models can be used to create synthetic data¹²⁹ or to select representative training samples,¹³⁰ which can train ML models that are both generalized and cost-effective. Unique neural network architectures that incorporate physical constraints, known as physics-informed neural networks, provide solutions to partial differential equations used in climate dynamics modeling.^131,132 In addition, embedding physical laws into the cost functions of neural networks, traditionally optimized by statistical measures such as cross-entropy or mean-square error, introduces a regularization effect and inherently discards physically implausible outputs.¹³³ The synergy between ML and physical modeling not only fortifies model credibility but also establishes a methodological evolution.

Future initiatives within the model-data-driven paradigm should concentrate on enhancing the scalability and integration of hybrid models across various scales and systems. This would include fine-tuning the interoperability between ML algorithms and process-based models to ensure seamless functionality in both regional and global-scale simulations. Such advancements could drastically improve the capability to simulate complex Earth system interactions and provide more accurate forecasts under changing climatic conditions.

This section has delved into the diverse paradigms and methodologies of geoscience, highlighting the multifaceted approaches to understanding our planet. The observational-hypothesis-driven paradigm forms the basis for empirical investigation, setting the stage for further inquiry, while the model-driven and data-driven approaches offer advanced simulation and in-depth analysis tools. In summary, the current landscape of research in geoscience has encountered limitations in effectively addressing complex global challenges.⁵⁹ There is a need for a transformative shift toward insights that integrate advanced AI techniques with geoscientific knowledge.¹²³ As geoscience continues to evolve, the interplay of these methodologies will be instrumental in driving forward our global efforts for environmental protection and sustainable development.¹⁰

AI-driven geoscience paradigms

Earth science research has undergone a transition from the observational-hypothesis-driven paradigm (see Figure 2) to a joint process-data-driven paradigm, which exhibits the characteristics of the “four Vs” of big data: volume, variety, velocity, and value.¹³⁴ Since the early 2010s, the performance of AI has improved dramatically⁷⁰ due to the availability of large-scale datasets, massive computer and storage hardware, and efficient distributed and parallel computing frameworks. The rise of AI has greatly accelerated the paradigm transition in geoscience research and driven various aspects of the application processes of big Earth data, from big Earth data collection and processing¹³⁵ to novel computational platforms,¹³⁶ hypothesis generation,¹³⁷ and geoscience prediction.¹³⁸ In this section, we discuss how AI can contribute to geoscience research in these aspects and the unprecedented opportunities it presents.

Figure 2.

AI-assisted observations, hypotheses, and predictions of geoscience

AI-assisted Earth observation data collection, processing, and representation

Data collection and analysis form the foundation of Earth science discoveries, aiming to capture, process, and represent complex Earth data to mine valuable information to understand complex Earth systems.^17,84 AI enhances and accelerates each stage of this process. AI accelerates and improves the efficiency of Earth observation data collection. The conventional satellite-to-ground data collection process typically includes multiple stages, requiring high time consumption and bandwidth.¹³⁹ Edge computing with AI on satellites allows real-time data processing and selective transmission to ground stations, significantly improving efficiency and reducing the need for manual corrections. Similar applications include real-time geographic information services for mobile terminals,¹⁴⁰ high-precision monitoring of ground stations,^141,142 and UAV-based agricultural remote sensing.¹⁴³ For example, Wang et al.¹⁴⁴ proposed a cloud-edge-end collaborative system for agricultural remote sensing, allowing AI to perform real-time data collection and processing on edge UAV devices. The processed data are then sent to the cloud, enhancing data transmission rates. In the future, integrated data collection and processing on edge sensors via AI¹⁴⁵ will be of great potential.

AI significantly contributes to data generation, completion, and enhancement. Earth observation data frequently encounter limitations in temporal, spatial, or spectral dimensions due to meteorological conditions, noise interference, and sensor issues, resulting in discontinuities across these dimensions.¹⁰⁹ Generative AI’s ability to process multi-modal data across time, space, and multiple spectra is essential for generating, completing, and enhancing geoscientific data.¹⁴⁶ For instance, Kadow et al.¹⁴⁷ developed an AI model using inpainting technology to reconstruct meteorological data, restoring the missing spatial pattern of the El Niño event from July 1877. Moreover, large diffusion models such as DiffusionSat¹⁴⁸ and CRS-Diff¹⁴⁹ are capable of performing integrated tasks and addressing the issue of limited remote sensing samples in specific spatiotemporal scenarios. SpectralGPT¹⁵⁰ captures spectral sequence patterns through multi-objective reconstruction, providing a foundation model with over 600 million parameters for various downstream tasks.

AI enhances the flexibility and effectiveness of data representations by introducing geometry and structure to model the complex interrelations within the data.^59,151 For example, graph networks^152,153 model directly underlying structures, facilitating the discovery of broader spatial correlation patterns in Earth science data. Self-supervised learning¹⁵⁴ allows capturing general features without relying on explicit labels. The Transformer architecture,¹⁵⁵ known for its powerful feature extraction and long-distance spatial dependency modeling capabilities, unifies data representations across various scenarios and modalities in Earth science. Recently, large AI models have revolutionized representation learning by facilitating deep interconnections between Earth science data to unearth new scientific discoveries. Examples include the single-modal large language model K2,¹⁵⁶ the meteorological time series graph network model GraphCast,¹⁵⁷ and the large multi-modal model SkySense, which integrates images, text, geographic coordinates, and site observations.¹⁵⁸ Exemplified by digital twin Earth, unified and universal large Earth science models have become a future trend.^150,159 Their embedding representations should not only consider the capabilities of multi-scale spatiotemporal data processing, multimodal data representation, and alignment with human understanding, but also provide a universal interface for decoders tailored to various downstream tasks, achieving comprehensive generalization across the domain.

AI promoting new computing tools or platforms for geoscience research

Numerous processes on and within the Earth are continuously monitored by various sensors globally, generating vast amounts of Earth data, with storage volumes exceeding 10 exabytes.^19,160 These sensors capture various states, fluxes, and intensities, capturing time/space-integrated data from satellite remote sensing, in situ observations, and atmospheric monitoring devices.¹⁹ Traditionally, geoscientific systems have required the integration of decentralized decoders tailored to specific tasks to compute and simulate the diverse and spatiotemporally varied streams of observational data. This approach often complicates data sharing and model connectivity. However, the emergence of new AI tools and large models is poised to revolutionize the computation and simulation paradigms of geoscience big data platforms.

First, large AI models are driving the innovative construction of big data platforms that offer robust multi-task processing capabilities and efficient data integration mechanisms. For example, AI-Earth¹⁶¹ introduced AI-Seg, a universal foundation model for object segmentation, capable of rapidly segmenting multi-source remote sensing images and extracting spatiotemporal change information. The Open Geospatial Engine¹⁶² incorporates LuojiaNet, a DL architecture tailored to geoscientific features, linking 55 downstream foundation models with 300 million parameters. This system also includes an embedded spatiotemporal knowledge graph to associate multimodal spatiotemporal data.

Second, AI agents, in collaboration with high-quality feedback from geoscience experts, can assist in solving complex geoscientific processes or problems. The “human-in-the-loop” process, which involves deep models and human experts, has proven effective in geoscientific data annotation with improved interpretation accuracy.¹⁶³ For instance, Li et al.¹⁶⁴ have integrated large conversational models into robots, allowing humans to issue commands to robots via language for complex action-planning tasks. This advancement in AI’s understanding of spatial intelligence is catalyzing robotic learning, approaching the goal of embodied intelligence.¹⁶⁵ Moreover, the deep integration of large AI models with drones, autonomous vehicles, and mobile monitoring devices on the surface or underground, coupled with satellite data, is facilitating more efficient and automated complex actions in geoscience, including spatiotemporal data collection, processing, and transmission.

Finally, the integration and assimilation of Earth’s big data through the digital twin Earth has ushered in a new era of experimentation and simulation in Earth science.¹⁶⁶ By integrating remote sensing data, in situ observations, experimental analyses, societal perceptions, simulations, and reanalysis, AI-based digital twin systems or platforms are capable of accurately simulating various complex Earth processes, spanning atmospheric, hydrological, urban, geological, and other domains.¹⁶⁷ Specifically, Earth digital twins, which integrate big Earth data and physics-based models within interactive computational frameworks, enable the monitoring and prediction of environmental changes and societal disruptions,¹⁶⁸ thereby driving a deeper understanding of Earth system processes and scientific cognition.

AI facilitating the generation and optimization of geoscientific hypotheses

Hypotheses are crucial research tools in Earth sciences, aiding scientists in comprehending the Earth system and its evolution through artificial observations and scientific conjectures.¹⁶⁹ For instance, Kepler¹⁷⁰ formulated the laws governing planetary motion based on extensive observations of stars and planets. Geoscientific hypotheses appear in various forms, including mathematical expressions, molecular formulas in geochemistry, and genetic variation laws in biology. Traditional methods for generating and validating hypotheses have predominantly relied on theoretical assumptions and logical deduction, as well as computational modeling and simulation,¹⁸ with limited ability to solve complex and nonlinear problems. In contrast, recent AI has learned patterns and rules in massive data through “guessing-and-verifying,” with intelligence gradually emerging.¹⁷¹ This evolution has led to significant breakthroughs in scientific endeavors, such as predicting protein structures,¹⁷² formally proving mathematical conjectures and theorems,¹⁷³ and simulating molecular dynamics in physics.¹⁷⁴ This “guess-and-verify” type of AI has greatly contributed to the paradigm shift of geoscientific hypotheses generation and validation.

AI is transforming the generation of geoscientific hypotheses from predefined methods to data-driven discovery. Traditionally, hydrologists modeled rainfall-runoff processes using physical conceptual models based on potential influencing factors,^175,176 which tend to be non-unique, subjective, and limited.¹⁷⁷ In contrast, AI treats multimodal data as inputs, enabling scientists to explore larger sets of hypotheses for more effective generation.^22,178 Furthermore, screening a high-quality hypothesis from the candidates is usually framed as an optimization problem.¹⁷⁹ AI prioritizes directions with higher values by maximizing reward signals for the candidate set, instead of using manually designed rules in the traditional approach.^180,181 For example, a multi-objective optimization framework was constructed to consider the impacts of hydropower capacity on five environmental factors (sedimentation, river connectivity, flow regulation, biodiversity, and greenhouse gases) in the Amazon basin.¹⁸² AI also enables selective screening of candidate information with desirable attributes from high-throughput experimental data, reducing the interference of redundant observations.¹⁸³ Another example is the optimization of discrete geo-hypotheses, where AI methods, such as variational autoencoders, map discrete symbolic representations into a differentiable latent space.¹⁸⁴ Process-based differentiable modeling⁶³ combines physical mechanisms and ML techniques, facilitating hypothesis testing and uncovering previously unrecognized correlations in Earth science.

AI holds the potential to significantly contribute to the verification of geoscientific hypotheses. Various hypotheses in geoscience, such as Wegener’s continental drift theory, Darwin’s biological evolution hypothesis, and the historical climate change conjecture, present major scientific challenges. Correspondingly, researchers have leveraged AI’s ability to model nonlinear complex systems to verify these hypotheses. For example, Stupp et al.¹⁸⁵ used coevolutionary ML to predict functionally relevant interactions between human genes, advancing the understanding of human coevolutionary processes. Kalra et al.¹⁸⁶ utilized artificial neural networks to model the complex association between global temperature and greenhouse gas concentrations. In addition, large AI models such as GraphCast¹⁵⁷ and PanGu¹⁸⁷ have revolutionized traditional weather forecasting methods and contributed to exploring the evolution of Earth’s climate over deep time. AI also challenges the findings of traditional physical models.¹⁷ For example, ML estimates of global carbon flux data have indicated that traditional climate models may have overestimated the response of vegetation, such as tropical rainforests and grasslands, to climate changes.¹⁸⁸ Data-driven carbon cycle estimates have also revealed potential mechanisms behind the enhanced seasonal cycle of atmospheric carbon dioxide concentration in high-latitude regions.¹⁸⁹

AI-driven solutions to geoscience inference and prediction

Geoscience prediction tools have undergone substantial evolution, improving our ability to comprehend complex Earth systems.¹⁹⁰ Initially, Galileo, Kepler, and others studied planets through experimental methods of observation and induction.¹⁹¹ Alfred Lothar Wegener studied the Earth’s plates through hypothesis and deduction.¹⁹² Subsequently, the simulation and modeling of complex phenomena, such as meteorological, hydrological, oceanic, and other physical processes, through physical computational models became the third paradigm of Earth science research. With the arrival of the big Earth data era and the continuous improvement of AI, scientists have made significant strides in spatiotemporal analysis.^151,193 The data-intensive research paradigm has become a mainstream.¹¹ Nowadays, large AI models have revolutionized the paradigms for geoscience inference and prediction,¹⁷ exhibiting strong abilities to mine hidden relationships within vast data and enhanced model inference and prediction accuracy.^194,195

AI allows for more comprehensive and efficient geoscience inference and prediction. To enhance geoscience inference, AI implements trustworthy attention-based models, enabling the extraction of spatial relationships across data from a global perspective.¹⁹⁶ Physical embedded neural networks leverage their powerful numerical approximation capabilities to reduce the computational complexity of high-order differential equations.¹⁹⁷ Spatiotemporal graph neural networks, which utilize the graph structures to accurately represent spatial relationships and factor correlations, enhance the reliability of reasoning.^195,198 In terms of prediction, AI incorporates a broad range of historical information and efficient modeling strategies, such as pre-training¹⁹⁹ and generative decoders,²⁰⁰ offering enhanced technical support for decision-making processes. With the deep integration of AI infrastructures (such as high-performance computing chips, storage media, rapid and lightweight large models) and edge sensors, real-time monitoring of the Earth’s environment contributes to enhancing the predictive capabilities for rapid disturbances such as geological disasters, climate anomalies, and emergencies.^91,154 Overall, AI facilitates more precise and reliable inference and prediction, reducing the computational complexity of high-order differential equations.^201,202

Society has witnessed many successes in this respect, although many challenges still exist. Weather prediction, a successful example in geoscience, has dramatically improved through integrating advanced AI models, increased computational power, and established observational systems with large amounts of data.¹²² Represented by PanGu¹⁸⁷ and GraphCast,¹⁵⁷ large AI models can accurately predict weather evolution on time scales ranging from several days to a month. However, challenges remain in seasonal weather forecasts, extreme event predictions (such as floods and wildfires), and long-term climate forecasts.^120,203 In the biosphere, Klemmer et al.²⁰⁴ trained a universal AI geolocation encoder to assist in monitoring biological population migration and number estimation. However, dominated by biologically mediated processes such as reproduction and migration, and influenced by seemingly random but intense disturbances such as earthquakes, landslides, and volcanic eruptions, predicting the dynamic changes and deep-time evolution of the biosphere remains difficult.^205,206,207 It is also essential to establish a comprehensive, integrated monitoring network in outer space, sky, surface land, and subsurface to provide more reliable data support.

Typical cases

AI, as a modern scientific research infrastructure that comprises rapidly evolving technologies, brings novel means to comprehend the Earth’s systems, including the atmosphere, lithosphere, hydrosphere, cryosphere, biosphere, and anthroposphere, as well as their interactions. By leveraging rapidly advancing technologies, AI accelerates and deepens our understanding of the Earth at a variety of spatial and temporal scales, advancing the achievement of sustainable development goals (see Figure 3). The uniqueness of geoscience, showcasing a considerable amount of subdisciplines, a vast quantity of geographic knowledge, an extensive collection of observational data and spatial dependence, spatial heterogeneity, and nonlinearities among geographical elements, has led to novel advancements in AI technology.

Figure 3.

Observation and simulation are the two main tools for understanding the Earth system

AI helps in the observation of the Earth system, assisting in the discovery of knowledge from data. Besides, AI also supports Earth system simulation, generating data from models and knowledge.

Atmosphere

Clouds, aerosols, and gases are three of the most important components in the atmosphere. They affect the solar radiation received by the Earth system and exert distinct radiative forcing on the energy budget, which in turn has a substantial influence on the weather and climate on a regional or global scale.^18,208,209 AI models the complex and nonlinear atmosphere system, predicting common surface and atmospheric variables, as well as enhancing our ability to retrieve atmospheric parameters with remarkable enhancement in the accuracy and granularity of atmospheric studies.^210,211

Atmospheric component detection and interactions

AI has revolutionized cloud identification, cloud type recognition, and even cloud dynamics prediction from satellites,²¹² It has notably improved the accuracy of retrieving cloud microphysical and cloud top parameters^213,214,215 and has provided cloud bottom information that traditional physical-based algorithms often fail to estimate.²¹⁶ These advancements enable the precise understanding of cloud formation in weather forecasting,²¹⁷ holding the promise of more accurate and efficient weather predictions.

In aerosol remote sensing, AI mainly contributes to improving the detection of aerosols,²¹⁸ building models to retrieve aerosol optical properties^219,220,221 and applying the aerosol products to wildfire detection, particulate matter (PM_2.5) monitoring, and other aerosol-related problems.²¹⁹ It is noteworthy that AI is becoming an irreplaceable tool to develop high spatiotemporal datasets of the aerosols originating from various emission sources that improve our understanding of the climatic, environmental, and health effects from the intricate composition of aerosols.^222,223,224

AI models, on the one hand, retrieve water vapor with high accuracy²¹⁶ and produce precipitation datasets with a high spatial and temporal resolution.²²⁵ On the other hand, AI techniques have been integrated with ground- and satellite-based observations to quantify and forecast air quality on a regional or global scale.²¹⁶ Many factors (meteorology, geography, emissions, vegetation, etc.) have been incorporated into the AI models to explore the complex non-linear relationships between satellite-based observations and the surface concentrations of various gaseous air pollutants, which provides insight into developing more efficient strategies to reduce the adverse health and societal effects of air pollution exposure.^{226,227,228,229}

By advancing cloud analysis, improving aerosol monitoring, and exploring the relationships between complex gases, AI significantly enhances researchers’ understanding of the dynamics of atmospheric components and captures their intricate interactions.

Solar radiation monitoring

The traditional radiative transfer (RT) model is a classic and widely used way to retrieve solar radiation. However, the forward RT simulation is a time-consuming process, which makes it inapplicable for direct use with satellite observations, particularly with geostationary satellite observations (the monitoring frequency in the order of minutes). Through the development of AI-based RT models in recent years, the computational efficiency of atmospheric RT has been greatly improved (by several orders of magnitude),²³⁰ which enables near real-time monitoring of solar radiation from satellites with high accuracy.²³¹

Weather forecasting and climate prediction

Mainstream AI-based global weather/climate forecast models predominantly concentrate on short- and medium-term predictions,¹⁵¹ such as Google DeepMind’s GraphCast,¹⁵⁷ Huawei Cloud’s Pangu-Weather,¹⁸⁷ Tsinghua University and China Meteorological Administration’s NowcastNet,²³² Alibaba’s SwinVRNN,²³³ Fudan University’s Fuxi,²³⁴ Shanghai’s AI Laboratory’s Fengwu,²³⁵ Microsoft and the University of Washington’s Deep Learning Weather Prediction,²³⁶ with exceptional capabilities in processing large datasets, performing real-time analysis, and predicting extreme weather events.²³⁷ AI-based global weather/climate forecasting models have high forecast timeliness and computational efficiency. Taking Pangu-Weather as an example, it predicts 7 days’ weather in only 10 s, 0.6 days earlier than the world’s leading weather forecasting system, the European Center for Medium-Range Weather Forecasts (ECMWF).¹⁸⁷ It is of great significance for extreme weather forecasting. Based on the weather forecast assessment of China’s National Ground Meteorological Stations in the first quarter of 2024, AI-based models such as Fuxi, GraphCast, and FourCastNet had higher accuracy in temperature and wind speed than traditional numerical predictions.

Atmospheric predictability revolution: From challenges to solutions

The atmosphere is an intricate and dynamic system, and myriad challenges originate from the subtle interaction among aerosols, clouds, gases, and radiations.²¹⁰ Predicting weather patterns and understanding climate change accurately are paramount in atmospheric science. However, achieving these goals poses significant challenges, including the need for faster and more precise weather forecasting and climate projections. Nowadays, AI models have emerged as a powerful tool for tackling these challenges and advancing solutions across a wide array of applications in atmospheric sciences.²³⁷ It significantly promotes the development of related monitoring and prediction platforms, which produce massive data and information with high spatiotemporal resolution and improved accuracy.^10,234 In the future, as AI continues to evolve and incorporate more spatial big data into its training, it will enhance the reliability and accuracy of weather and climate forecasts further. Consequently, it may even lead to the eventual replacement of traditional physics-based models with AI-driven approaches. In addition, AI will play an imperative role in constructing automated monitoring and warning systems for the atmosphere environment, enabling timely issuance of alerts and recommendations. In essence, AI’s application in atmospheric science transcends traditional methods, providing innovative solutions to long-standing challenges. Its integration into timely and accurate monitoring and prediction systems not only advances our understanding of atmospheric processes but also empowers us to make well-informed decisions.

Lithosphere

Solid Earth science, aimed at comprehending the structure, materials, and dynamics of the Earth’s interior, geological processes, and the evolutionary history of the Earth,²³⁸ receives unprecedented opportunities from AI,¹⁰ with dramatic developments in geological hazard monitoring and prediction, rock feature analysis, geological exploration, geological model construction, and analysis of soil characteristics.^10,239

Geological exploration and hazard prediction

AI approaches have made significant strides in their application to geological exploration, such as petroleum and natural gas exploration,²⁴⁰ geophysical imaging,²⁴¹ as well as the processing of seismic,²⁴² magnetotelluric,²⁴³ and gravity data,²⁴⁴ enabling sophisticated analysis and interpretation. Techniques such as denoising, phase-picking, and weak signal enhancement can reduce human errors in the exploration process, enhance the quality of exploration data, and accelerate exploration time.²⁴⁵ By harnessing the power of AI, geoscientists can unlock new frontiers in exploration efficiency, accuracy, and cost-effectiveness, ultimately shaping the future of resource exploration and sustainability.

AI technology provides powerful tools in facilitating earthquake monitoring and prediction, including detection and phase identification,²⁴⁶ early warning,²⁴⁷ motion prediction,²⁴⁸ as well as forecasting magnitudes, scales, and timing,²⁴⁹ also in assessment of landslide susceptibility,²⁵⁰ supporting the mitigation of risks. AI has also demonstrated great potential in the volcanic prediction process.²⁵¹ Although AI encountered grand challenges in operational earthquake prediction, forecasting of fault zone stress, and the occurrence of chained natural hazards attributed to their highly coupled and strongly non-linear dynamics,²⁵² it has exhibited tremendous progress in recent years.²⁵³

Rock physics analysis

AI methods can be utilized for the analysis and classification of rock samples,²⁵⁴ automatically identifying rock types, compositions, and physical characteristics, thus expediting the analysis of rock samples and providing more detailed information about rock features, so as to aid geologists in better understanding geological history and rock evolution.²⁵⁵ Recent evidence demonstrates that AI has successfully solved various problems in rock mechanics, outperforming conventional empirical or statistical methods.²⁵⁴ By using AI approaches, deeper insights can be gained for more accurate geological interpretations and predictive models.

Geological modeling

As the emerging paradigm of science and technology research, AI is modulating the world in a variety of science realms, including geology. AI is transforming the measures geologists analyze data and understand the mechanisms of deep Earth. During the construction of geological models, AI is capable of integrating vast amounts of geological data from various disciplines and fields, ranging from geophysics and geochemistry to hydrology and tectonics. This multidisciplinary approach generates predictive models that assist scientists in better comprehending subsurface structures, stratigraphic forms, and groundwater flow. AI-driven predictive modeling helps geologists to efficiently and accurately identify patterns and trends that were difficult to detect early on.^256,257,258 AI’s integration into deep Earth modeling enables geologists to identify previously unrecognized geological features and phenomena, thereby advancing our understanding of the deep Earth.

Soil characteristics monitoring

AI boosts new developments in soil monitoring, offering a holistic and data-driven approach to soil monitoring and management.²⁵⁹ AI-driven sensors and monitoring systems enable continuous and high-resolution monitoring of soil conditions, providing valuable insights into soil health and dynamics, including essential soil parameters such as moisture,²⁵⁹ temperature,²⁶⁰ and texture.²⁶¹ By analyzing multispectral and hyperspectral imagery, AI models can accurately map soil types, nutrient levels, and organic matter content across large spatial scales. This new real-time ability assists farmers in making informed decisions, thereby improving farmland utilization efficiency and agricultural production quality.²⁶²

Deep-time and deep-Earth discoveries: Scale and accuracy

To date, large AI models constitute the most cutting-edge and wisdom-intensive research regime; the integration of AI has indeed ushered in a new era of exploration and understanding. However, as we delve deeper into the complexities of lithospheric processes, it becomes apparent that simply scaling up large AI models without due consideration for their ability to accurately capture and resolve scientific intricacies may lead to deviations from fundamental physical laws and characteristics. While large AI models boast impressive computational power, their efficacy in accurately describing lithospheric phenomena may be limited by the uncertainties inherent to input labels and data. Therefore, a shift toward the development of numerous and accurate small-scale domain-oriented models tailored to specific scientific problems or application fields is warranted. In addition, the integration of high-quality observation, monitoring, and experimental data with completeness is crucial for training and validating AI models in lithospheric studies.²⁶³ Synthetic data derived from massive-scale numerical simulations can further enhance the robustness and generalizability of AI models. Essentially, it may be a reliable and feasible measure to promote the revolutionary engagement of AI in deep-time and deep-Earth discoveries.

Hydrosphere

The hydrosphere is the sum of all water, including atmospheric, land surface, oceanic, and underground water reserved on Earth.²⁶⁴ AI addresses a wide range of applications in the hydrosphere that allow (but are not limited to) better modeling and estimation of precipitation, soil moisture, evapotranspiration, streamflow, water storage, ocean currents, and ocean salinity, by simulating the complex input-output relationships inherent to nonlinear hydrological processes, thus improving the accuracy of hydrological model simulations and remote sensing retrievals.^265,266,267

Land surface water balance

AI benefits the closing of the land surface water balance by accounting for the individual surface water flux components (precipitation, evapotranspiration, streamflow) and expanding the mapping capabilities of key state variables (such as soil moisture). AI can improve the estimation and forecasting accuracy of precipitation and help better understand the causes of extreme rainfall.²⁶⁸ For example, generative adversarial networks have been used for precipitation nowcasting and proved to be of high reliability.²⁶⁹ The multi-layer perceptron model, integrating geostationary satellite infrared data and passive microwave-based retrievals, yields precise precipitation estimates.²⁷⁰ Probabilistic weather models such as deep neural networks (i.e., MetNet-2) forecast precipitation with exceptional resolution, up to 12 h ahead.²⁷¹ Moreover, AI empowers the generation of precipitation data with unparalleled precision, spatiotemporal resolution, and spatial coverage, enhancing our understanding of precipitation dynamics.²⁷² In addition, AI methods analyze large-scale circulation patterns associated with US Midwest extreme precipitation to better understand the physical causes of changing extremes.²⁷³ Despite the many successful cases of AI application, acquisition of high-quality and continuous atmospheric data is still challenging due to sensor limitations, and the implementation of hybrid models appears as an effective solution.

AI-based approaches have been extensively employed to estimate evapotranspiration, one of the most important components of the hydrological cycle. That is crucial for estimating irrigation water requirements, hydrological processes, and assessing agricultural systems at both regional²⁷⁴ and global scales.²⁷⁵ Site-scale evapotranspiration observations can be upscaled to the regional scale using AI-based methods,²⁷⁶ thus overcoming the limited spatial and temporal coverage of in situ observations. The ability of AI to forecast evapotranspiration is also highlighted in a recent study.²⁷⁷ These forecasts play a crucial role in agricultural planning and drought monitoring, contributing to improved resilience and sustainability in water management practices. A novel research direction is to estimate evapotranspiration at high resolution through the construction of hybrid models,²⁷⁸ which combine the physical consistency and interpretability of physical models with the data-driven formulations of AI-based models, thereby revealing processes that are insufficiently understood. This interdisciplinary approach holds the potential to uncover the underlying mechanisms and diversity of evapotranspiration, thereby enabling more robust and insightful assessments of water cycle dynamics.

Streamflow, as a key aspect of sustainable water resource planning and management, can be estimated in real time²⁷⁹ or forecasted at lead times of 1–7 days²⁸⁰ AI-based approaches, successfully used in streamflow regionalization,^281,282 can help to reduce modeling errors in process-based hydrologic models to improve the accuracy of simulations, since process-based and AI approaches can complement each other with respect to their inherent strengths and limitations.²⁸³ Deep neural networks enable the accurate identification of spatial distribution and morphological features of water bodies,²⁸⁴ understanding river evolution, and forecasting river dynamics,²⁸⁵ performing water quality analyses on the catchment scale.²⁸⁶ Another significant contribution of AI is the creation of global water quality databases due to its powerful learning and data fusion capabilities, such as the Global Streamflow Indices and Metadata Archive,²⁸⁷ global river discharge reanalysis,²⁸⁸ Global River Chemistry Database,²⁸⁹ and Global River Water Quality Archive.²⁹⁰ The integration of AI into streamflow estimation, forecasting, and water quality analysis offers transformative opportunities for strengthening our understanding of hydrological processes toward more sustainable and resilient water systems.

Soil moisture acts as a fundamental boundary condition in terrestrial hydrology.^178,291 The integration of AI-based models into soil moisture mapping significantly advances our ability to accurately retrieve, downscale, and predict soil moisture dynamics across different spatial and temporal scales. By using AI-based models, soil moisture retrievals are obtained from the passive-only²⁹² and synergistic active-passive microwave observations^293,294 with improved accuracy and temporal resolution, which is challenging for traditional algorithms to separate and interpret the desired information accurately. AI techniques downscale soil moisture from coarse spatial resolution to fine resolution,^{295,296,297,298,299} also establish long-term global daily surface soil moisture datasets from multi-frequency radiometers (AMSR-E/2 and FY-3 series) by transferring the Soil Moisture Active Passive L-band observations, offering extended records vital for climate monitoring and hydrological research.^300,301 Moreover, with the help of AI algorithms, soil moisture can be predicted at deeper depth (e.g., root zone) from surface data^302,303 and in a seamless and efficient manner through AI-based data assimilation techniques.^304,305

AI provides a potential solution for avoiding closure errors by enhancing the estimation and prediction of individual water fluxes and state variables. It addresses challenges related to integrating diverse data sources to produce cohesive models, achieving fine-scale spatial and temporal resolution, and understanding the nonlinear nature of hydrological processes.

Terrestrial water storage

AI plays an important role in improving the spatial and temporal continuity and resolution of terrestrial water storage. AI approaches have been instrumental in reconstructing continuous total water storage, by filling the data gap between the Gravity Recovery and the Climate Experiment (GRACE) satellite mission and its successor, GRACE-FO.³⁰⁶ Similarly, AI-based models, such as the GTWS-MLrec, have been developed to reconstruct terrestrial water storage estimates spanning several decades from 1949 to 2022, using a set of ML models with a large number of predictors.³⁰⁷ AI was used to capture complex spatiotemporal patterns in water storage dynamics, facilitating comprehensive analyses of hydrological trends and variability over extended periods. In addition, AI-based approaches have been deployed to map soil water storage in Ghana at high spatial and temporal resolutions, facilitating the identification of areas with stable water availability for improved crop production and guiding drought adaptation strategies.³⁰⁸ Moreover, GRACE-derived terrestrial water storage anomalies are downscaled to 10 km spatial resolution by using a convolutional long short-term memory neural network³⁰⁹ in Iran and convolutional neural network-based approaches in Canada.³¹⁰ In essence, AI provides a new capability to overcome data gaps, improve spatial resolution, and enhance the continuity of water storage observations, ultimately contributing to more effective water resource management.

Ocean currents and salinity

Ocean currents and salinity are crucial for understanding global climate systems, marine ecosystems, and coastal environments. Ocean currents reflect the movement of ocean water and drive the distribution of heat, nutrients, and salinity, influencing weather patterns, climate regulation, and marine biodiversity. AI methods significantly improve the estimation and forecasting of ocean currents by enhancing computational efficiency and accuracy.³¹¹ Traditional methods often struggle with the complexity and volume of oceanographic data. AI models, such as those integrating sea surface height, temperature, and wind stress simulated from the ocean general circulation model, can accurately predict the ocean currents over most of the global ocean,³¹² and successfully forecast the velocity. For structures of the loop current system,³¹³ AI techniques such as LSTM recurrent neural networks and the Transformer also enable real-time in situ prediction of ocean currents at any location, and overcome the problem of excessive computational complexity in traditional regional physics-based prediction models.³¹⁴

AI-based approaches, such as deep neural networks, generative adversarial networks, random forests, support vector regression, and multi-layer perceptrons promote the convenient and fast estimation of ocean salinity, from the Aquarius,³¹⁵ SMAP,³¹⁶ and the Geostationary Ocean Color Imager-II satellites.^317,318 With the aid of AI-based methods, the ocean general circulation model (e.g., Hybrid Coordinate Ocean Model) is also able to achieve more reliable estimates of sea surface salinity.³¹⁹ AI has demonstrated strong capabilities to reconstruct the high-precision and high-resolution three-dimensional (3D) ocean subsurface salinity on a daily scale in 12 depth levels (from 2 to 200 m) only relying on the ocean 3D temperature data.³¹⁵ This is because AI, particularly the DL models, have flexible structures and can extract potential complex mappings of data by stacking only multiple nonlinear layers.

Extreme hydrological events: Pioneering solutions

The changing dynamics of global climate present two concerning trends in the hydrosphere: alterations in water circulation patterns and the increasing frequency and intensity of extreme hydrological events.^111,320 In response to these challenges, it becomes imperative to strengthen monitoring efforts, enhance forecasting capabilities, and improve decision-making efficiency. AI provides important tools for monitoring, understanding, and forecasting extreme hydrological events such as drought, rainstorm, and flood.¹²² AI can integrate a large amount of data from various sources (e.g., satellites, meteorological stations, and other sensors) to provide more comprehensive and accurate monitoring results of extreme hydrological events.³²¹

For example, for extreme events, Earth observation data and ML can significantly mitigate the scarce hydrological data. Satellite-based technologies, which encompass a wide array of sensors operating across different regions of the electromagnetic spectrum—such as visible, thermal, and microwave domains—offer considerable potential. Advanced sensors, including synthetic aperture radar (SAR), satellite-based precipitation measurements, and gravity measurements, are emerging as transformative tools for the forecasting and monitoring of extreme events.³²² Concurrently, the robustness and transferability of ML techniques are proving instrumental in predicting floods in ungauged river basins.¹²²

Moreover, AI can analyze and learn from historical data and meteorological forcings (such as precipitation and temperature), and identify the interactions between different environmental factors, and thus help understand the causes and patterns of extreme hydrological events.³²³ Furthermore, by using AI, short-term forecast of hydrological events can be made based on real-time hydrological data, providing timely support for emergency response. Short-term flood forecasting, which spans from a few hours to several weeks, predominantly utilizes meteorological forecasts to enhance model prediction performance and ensure physical consistency. For example, Xu et al.³²⁴ have summarized numerous hydrological forecast models in this context. The prevailing methods for short-term flood forecasting integrate meteorological inputs (such as precipitation and temperature) with optional historical data to predict runoff or flooding events.

Meanwhile, combining meteorological and hydrological models, AI can forecast the long-term trend of extreme hydrological events, helping decision-makers to make long-term plans.³²⁵ Long-term forecasting of extreme events, which includes sub-seasonal, annual, and decadal outlooks, remains a significant challenge due to inherent data and model uncertainties. Currently, hybrid learning approaches³²⁴ that combine physical modeling with ML are being employed to reduce model uncertainties and mitigate the reliance of data-driven models on extensive data inputs. In addition, uncertainties in data (such as precipitation) can be addressed by integrating low-latency satellite observation data with reliable climate prediction models. In summary, AI has brought new opportunities for hydrological cycle research to better understand and cope with extreme hydrological events. With the rapid development of computer technology and the emergence of new interpretable AI methods, the role of AI in the hydrosphere (particularly in extreme hydrological events) will become more prominent in the future.

Cryosphere

The cryosphere refers to frozen components of the Earth system,³²⁶ overlapping with the atmosphere, the hydrosphere, and the lithosphere over vast areas, exhibiting a sensitive response and holding a significant impact on climate change.^327,328 Numerous scholars focused on developing AI methods for addressing the challenging geoscientific questions in cryosphere research, such as the AI for Cold Regions, bringing new perspectives and innovative solutions in element classification and automatic mapping, feature spotting, physical properties retrieval, and interpretation of the cryosphere changes.³²⁹

Cryosphere element identification

AI overcomes ambiguity in the cryosphere element identification caused by feature similarity, superseding manual interpretation, and limited empirical approaches. One notable application of AI is that it enhances our comprehension of the spatiotemporal distribution of the cryosphere by better classifying its elements, such as distinguishing ice cover types,³³⁰ especially debris-covered glaciers,³³¹ which were difficult for band ratios/indices. AI can overcome inherent complexities to generate high-resolution maps of permafrost, a critical component of the cryosphere.³³² Kuter et al.³³³ applied artificial neural networks to estimate areal snow cover extent with high accuracy, thus able to provide timely and reliable information on snow cover dynamics. Convolutional neural networks have proven effective in classifying sea ice types with higher accuracy and less sensitivity to noise in SAR images.³³⁴ Coincidentally, AI achieves automatic and reliable iceberg detection in different environmental conditions and improves understanding of iceberg dynamics in polar regions.³³⁵ In essence, AI plays a significant role in spearheading our understanding of the cryosphere by overcoming traditional limitations in element identification.

Feature spotting

In addition to classifying the cryosphere elements, AI aids in the identification of specific features of these elements that were previously challenging to detect. Specifically, AI has advanced the identification of wet and dry snow, especially in vegetated and mountainous areas where traditional methods struggle to differentiate between snow types.³³⁶ AI enabled robust and automated detection of snow avalanches for enhancing safety measures in mountainous regions.³³⁷ In glaciological research, AI has been utilized to map glacier calving margins³³⁸ as well as glacier termini³³⁹ toward comprehensive assessments of glacier mass loss. Qayyum et al. developed a DL-based glacial lake extraction method with noteworthy benefits in monitoring glacial lakes, a key indicator of potential glacial lake outburst floods.³⁴⁰ In permafrost research, ML performed analysis on the distribution of retrogressive thaw slumps³⁴¹ and extraction of ice-wedge polygons.³⁴² Beyond that, AI has been used to improve the quantification of sea ice surface coverage types, and also to extract Antarctic ice shelf fronts from Sentinel-1 Imagery³⁴³ and to classify ice crystal habitats more precisely than traditional methods.³⁴⁴ Therefore, AI plays a key role in promoting frontiers in cryospheric research by enabling the detection and characterization of specific features within cryospheric elements.

Properties retrieval

Different from traditional and complex physical models, AI enables simplified yet accurate property retrieval by modeling multivariate nonlinear relationships between cryospheric element parameters and image characteristics. This paradigm shift has led to significant advancements in understanding cryospheric processes. For example, AI improves the retrieval accuracy of the cryosphere properties in coalition with conventional algorithms, such as retrieval of snow depth³⁴⁵ and estimates of snow water equivalent,^346,347 providing new insights to hydrological processes in cold regions. AI helped to solve the problem of detecting each internal ice layer uniquely to estimate their thickness accurately, thus providing crucial insights for assessing the contributions of ice sheets to sea level rise.³⁴⁸ In permafrost research, AI has been applied to estimate mean annual ground temperature and active layer thickness and to estimate the thaw depth variations at seasonal scale.³⁴⁹ AI has achieved better performance in Arctic sea ice thickness estimation, a key indicator of Arctic climate change.³³² In addition, AI has helped to reconstruct the winter glacier mass balance, a quantitative expression of glacier volume change through time, filling the gap in ground observations and providing valuable insights into long-term glacier volume changes.³⁵⁰ Therefore, AI has led the way in streamlined and accurate cryosphere property retrieval.

Trend projection

AI significantly improves trend forecasting across diverse and complex conditions by developing sophisticated models that enhance spatiotemporal scope and precision. AI facilitates the investigation of historical cryospheric changes of possible trends, such as improving the prediction sensitivity of arsenic or manganese in groundwater and identifying trends that may not be apparent through traditional methods alone.³⁵¹ Similarly, AI was used to model the future responses of permafrost to climatic changes,³⁵² including permafrost degradation trends,³³¹ overcoming limitations of environmental conditions. In addition, AI was applied to estimate snow avalanche hazards for a better prediction of occurrence and magnitude.³⁵³ AI also advanced the range of accurate sea ice forecast.³⁵⁴ Regarding iceberg research, AI has been used to estimate the surface area and masses of icebergs,³³⁴ which has operational difficulties in large-scale monitoring by observational and remote sensing methods. Therefore, AI-driven approaches significantly propel trend forecasting and predictive modeling within the cryosphere, providing valuable insights into historical changes, future projections, and operational challenges.

Cryospheric water storage dynamics and sea level rise

The cryosphere, a critical component of Earth’s climatic system, is rapidly diminishing due to the effects of global warming. This trend is particularly evident in glaciers, including the massive Greenland and Antarctic ice sheets, which are experiencing accelerated mass loss. Moreover, sea ice coverage and snow extent are decreasing, while permafrost is undergoing significant degradation. This shrinking cryosphere is directly contributing to rising sea levels, posing imminent and long-term threats to low-lying coastal areas and small island nations. In addition, in mountainous regions and high plateaus, the reduction of cryospheric elements is causing fluctuations in river runoff, exacerbating water scarcity and increasing the risk of flooding in vulnerable areas. Cryospheric elements, such as glaciers, snowpacks, permafrost, sea ice, and ice caps, possess 3D or stereo characteristics. Traditional Earth observation methods often provide surface properties or limited-depth information, hindering comprehensive assessments of cryospheric elements. AI presents an opportunity to enhance our understanding of the 3D properties of cryospheric elements. For instance, AI can provide improved models of the active layer in permafrost and quantitatively assess the future conditions of permafrost.³⁴⁹ AI-enhanced algorithms can better align with field data of snow depth.³⁴⁵ Similarly, AI can improve sea ice thickness estimation algorithms to predict changes.³⁵⁵ Utilizing AI for assessing mass balance from ice sheet volumes has the ability to estimate its contribution in sea-level rise, offering a new methodology of climate change studies.³⁵⁶ In addition, AI has improved the precision of identifying each internal ice layer thickness in radar images, overcoming the limitations of traditional feature detection.³⁴⁸ By combining Earth observation technologies, physical modeling, and AI techniques, researchers can delve deeper into the interior of the cryosphere, gaining crucial insights into its formation, evolution, and distribution. This integrated approach not only improves our understanding of cryospheric stereoscopic characteristics, but also enhances climate change research, particularly concerning cryosphere melting and its implications for sea-level rise.

Biosphere

Recent advances in satellites and aerial missions have led to the accumulation of ecological data streams, leading to the development trend of ML and DL models to advance our knowledge of the biosphere, including ecological parameter inversion and characteristics mapping.^{113,357,358,359,360,361,362}

Vegetation properties mapping

Utilizing automatic learning of relationships between hundreds of bands and target variables, ML techniques such as decision trees, neural networks, and support vector machines have demonstrated exceptional efficiency in mapping vegetation structural and biochemical properties, encompassing leaf chlorophyll content, vegetation nitrogen, canopy cover, and leaf area index. In addition, ML algorithms play a crucial role in upscaling carbon fluxes (e.g., gross primary production, net ecosystem exchange, and ecosystem respiration) at regional and global scales.

Extracting vegetation variables is essential for evaluating how vegetation responds dynamically to fluctuating environmental conditions.¹² Utilizing automatic learning of relationships between hundreds of spectral bands and target variables, ML techniques such as decision trees, neural networks, and support vector machines have displayed outstanding performance in mapping vegetation structural and biochemical properties. These advanced algorithms effectively quantify parameters such as leaf chlorophyll content,^363,364 vegetation nitrogen,³⁶⁵ canopy cover,^364,366,367 and leaf area index,^368,369,370 showcasing a substantial improvement over traditional empirical methods. These AI-driven models offer not only increased accuracy but also remarkable scalability and adaptability across different environmental conditions.^371,372,373

Ecological parameter retrieval

In addition to mapping the vegetation properties and carbon fluxes, AI has advanced the precise identification of critical ecological parameters that were previously challenging to detect quickly and widely in terms of fine scale. Specifically, AI has achieved better performance in 3D structural parameters of forests such as leaf morphology,⁹¹ tree height,³⁷⁴ tree diameter at breast height,³⁷⁵ and ground vegetation canopy size.³⁹ Similarly, AI techniques have also been employed in marine plankton structure.³⁷⁶ In addition, AI helped to solve the problem of detecting and monitoring ecological disturbance.³⁷⁰ Previous attempts have been based on laborious and complex handcrafted extraction of image features, but in recent years it has been shown that sophisticated convolutional neural networks can learn to extract relevant features automatically,³⁷⁷ without human intervention. Automated image interpretation with convolutional neural networks performs very well for monitoring forest diseases and pests, close to human performance, and that makes professional field campaigns less costly.²²⁵ In agricultural monitoring research, AI promotes the identification of malnourished crops, thereby assisting in the precise management of farmland.³⁷⁸ Furthermore, AI has made significant progress in fine-scale geographic information simulation and prediction. Specifically, the rapid development of DL has notably enhanced the precision of urban characteristics simulating refined features more precisely than traditional methods.³⁷⁹ AI also advanced the refined simulation of surface temperature and addresses the previously unresolved issue of fine simulation of extreme urban heat island effects.³⁸⁰

Fine-scale ecology analysis

On even finer scales, AI has achieved better performance in identifying ecological elements, promoting quantitative research on micro-ecosystems. In the research of diagnosing insects, AI techniques have reached 97% accuracy and outperformed a leading taxonomic expert.^381,382 For the identification and classification of vegetation pollen, DL technology has achieved automated pollen analysis methods,³⁸³ which greatly solves the labor cost of labor-intensive pollen analysis in the past and significantly improves analysis efficiency. In addition, as a crucial means of extracting geographic information, classification technology has evolved further with the aid of AI foundations.^367,384 Currently, DL exhibits significant advantages in urban canopy detection³⁷⁰ and tree species classification,³⁰⁹ among others. By training with a large amount of data, DL-based models can achieve good prediction results for complex phenomena, such as crop element classification³⁸⁵ and high-precision urban land element classification.³⁸⁶ Simultaneously, existing experimental results demonstrate the superiority of the proposed AI model for both road detection and centerline extraction tasks.³⁸⁷ Meanwhile, the integration of DL with high-resolution remote sensing images enables the refinement of ground feature statistics, which has advantages for separating and interpreting the desired information accurately over traditional remote sensing algorithms. For example, the U-Net neural network was employed to count trees in Africa,^377,388 which has operational difficulties in large-scale monitoring by observational and remote sensing methods. Overall, there is little doubt that there are many opportunities for trait-based ecology to benefit from the integration of computer vision and AI.

Global carbon budget

Accurate assessment of carbon dioxide uptakes and emissions of the terrestrial biosphere is critical to better understand the global carbon cycle, support the development of climate policies, and project future climate change.^93,384 AI plays an extremely important role in integrating satellite remote sensing and carbon fluxes from in situ observations to achieve high-precision, high-resolution scientific data on carbon fluxes of terrestrial ecosystems at regional and global scales.^228,389 For example, ML has been applied to estimated global plant gross primary production, net ecosystem exchange, ecosystem respiration, and soil respiration by integrating multi-source remote sensing data (i.e., various temperature, moisture, and plant production-related remote sensing products) and carbon fluxes data from ground observations.^390,391 The comparative advantages of AI over traditional methods are primarily due to its ability to effectively incorporate nonlinear relationships between remote sensing data and carbon fluxes. Thus, AI could assist the global carbon budget by providing more accurate and higher-resolution global plant production and ecosystem respiration detection.

Other domains

In addition to the aforementioned five spheres, AI is also significantly involved in other domains such as anthroposphere and inter-/cross-spheres, along with the engagement in sustainable development, opening new perspectives for analysis, interpretation, and fostering a more balanced relationship between human society and Earth’s systems.³⁹²

Human activities understanding

Al plays a crucial role in comprehending and managing Earth’s complex systems and environments, serving as a formidable toolset to glean insights, anticipate trends, and devise effective strategies for sustainable development and resource management. AI’s multifaceted applications are particularly evident in its utilization by scientists for the analysis of real-time video streams derived from surveillance cameras and satellite imagery. This analytical prowess enables behavior analysis and large-scale monitoring of human activities, thereby offering invaluable insights into lifestyle patterns and social dynamics.³⁹³ By harnessing AI-driven analytics, researchers can discern nuanced behavioral patterns, track movement trends, and identify emergent phenomena, facilitating a deeper understanding of human interactions with the environment and informing evidence-based decision-making processes.

Furthermore, AI serves as a cornerstone in the realm of urban development assessment, facilitating comprehensive analyses including diverse facets such as urban expansion, infrastructure changes, etc.³⁹⁴ Leveraging AI-powered algorithms, urban planners and policymakers can assess the spatial dynamics of urban growth, anticipate infrastructure demands, optimize transportation networks, and devise sustainable land-use strategies. By amalgamating geospatial data with advanced analytical techniques, AI empowers stakeholders to make informed decisions aimed at fostering resilient, inclusive, and environmentally sustainable urban environments.

In tandem with its applications in physical environment monitoring, AI assumes a pivotal role in unraveling the intricacies of human behavior and preferences in the digital sphere. Social media analysis augmented by AI algorithms offers a potent lens through which online behavior and preferences can be discerned, thereby facilitating targeted advertising, personalized recommendations, and sentiment analysis.^146,395 By scrutinizing vast troves of user-generated content, AI-driven analytics can unveil latent trends, identify influencers, and gauge public sentiment, thereby enabling businesses and marketers to tailor their strategies to resonate with their target audience effectively.

In a word, AI’s integration into Earth’s complex systems and environments represents a paradigm shift in our ability to comprehend, monitor, and manage the multifaceted interplay between human activities and the natural world. By harnessing AI-driven analytics, researchers, policymakers, and businesses can unlock unprecedented insights, foster informed decision-making, and pave the way for a more sustainable and resilient future. However, it is imperative to acknowledge and address the ethical, privacy, and equity considerations inherent in the deployment of AI-powered systems, ensuring that these technologies are leveraged responsibly to serve the collective interests of humanity.

Spheres’ interactions

AI has emerged as a powerful tool for capturing inter-layer relationships and enhancing simulations of biogeochemical cycles.³⁹³ By leveraging AI techniques, such as DL, researchers can gain deeper insights into Earth’s historical evolution and phenomena such as the snowball Earth event.¹⁶⁸ One notable advantage of AI in this context is its ability to improve computational efficiency³⁹⁶ and parameter optimization,³⁹⁷ thereby facilitating more accurate and robust simulations. In addition, AI aids in predicting matter exchange patterns and developing effective adaptation strategies to manage environmental changes.

Furthermore, AI contributes to refining our understanding of Earth’s energy budget by integrating DL algorithms with remote sensing applications and incorporating biogeophysical feedback into models of the water cycle.^398,399 This interdisciplinary approach enables researchers to assess land surface changes and their impacts on energy budgets. Moreover, AI helps address the risks associated with over-parameterization in models, ensuring that simulations remain realistic and reliable. By identifying critical thresholds that trigger extreme events in Earth’s systems, AI plays a crucial role in various applications, including volcano alerts,⁴⁰⁰ groundwater mapping,⁴⁰¹ and studying climate-vegetation relationships.⁴⁰² This capability is crucial for improving early warning systems and mitigating the impacts of natural disasters on human populations and ecosystems.

The potential of AI extends beyond individual applications to regulating inter-layer dynamics and foreseeing thresholds that transform interactions at different scales. This proactive approach to exploring Earth’s systems and managing its resources holds promise for sustainable Earth management. By leveraging AI technologies, researchers in geoscience can better anticipate and respond to environmental challenges, paving the way for more effective conservation efforts and informed policy decisions.

In conclusion, AI offers significant opportunities for advancing our understanding of Earth’s complex systems and enhancing our ability to manage and protect the planet. By harnessing AI’s capabilities in capturing inter-layer relationships, optimizing simulations, and identifying critical thresholds, researchers can contribute to proactive exploration and sustainable Earth management. However, realizing this potential requires continued interdisciplinary collaboration and the responsible deployment of AI technologies in geoscience research and environmental conservation efforts.

Sustainable development goals

The United Nations’ 2030 Agenda outlines 17 interlinked goals that are set to solve development issues in economic, social, and environmental dimensions and realize sustainable development by 2030.⁴⁰³ These goals interrelate closely with the Earth’s spheres (lithosphere, hydrosphere, atmosphere, biosphere, and anthroposphere), aiming to ensure their equilibrium for human well-being and environmental sustainability. The appeal of leveraging AI to advance social benefits and achieve sustainable development goals (SDGs) has captured the attention of numerous practitioners and researchers.^404,405 For instance, in exploring the 169 targets outlined for the 17 goals, Vinuesa et al.⁴⁰⁶ demonstrated that AI serves as an enabler for 134 targets while acting as an inhibitor for 59 targets. Gupta et al.⁴⁰⁷ and Nasir et al.⁴⁰⁸ delved into discussions about the implications of AI on the SDGs at the indicator level.

• (1)Economic sustainable development goals. The technological benefits facilitated by AI also hold the potential to positively impact the attainment of several SDGs within the Economy group (SDGs 8, 9, 10, 11, and 12). Acemoglu and Restrepo indicate a net positive effect of AI-enabled technologies linked to increased productivity, highlighting potential negative consequences, particularly heightened inequalities.⁴⁰⁹ If future markets heavily rely on data analysis and these resources are not equitably available in low- and middle-income countries, it could significantly widen the economic gap, exacerbating inequality even within nations.⁴¹⁰
• (2)Social sustainable development goals. For SDGs 1, 2, 3, 4, 5, 7, 16, and 17, in the social group, AI acts as an enabler for all the targets by supporting the provision of food, health, water, and energy services to the population, enhancing poverty mapping, identifying vulnerable populations, and optimizing resource allocation.^411,412 AI-based applications, including smart traffic management, waste management, and energy-efficient infrastructure, etc., contribute to developing sustainable and resilient urban developments.^413,414
• (3)Environmental sustainable development goals. The potential of AI extends to the analysis of extensive interconnected databases for collaborative initiatives aimed at environmental preservation (SDGs 6, 13, 14, and 15).⁴¹¹ AI aids in water management through predictive analytics, monitoring water quality, and optimizing distribution networks.⁴¹⁵ AI is also poised to create low-carbon energy systems with the integration of renewable energy and essential components in climate 800 change, such as detecting the forest changes in satellite images to support habitat monitoring and decision-making.^416,417

Large models in geoscience

In this section, our principal objective is to elucidate the most recent developments associated with large models in geoscience,⁴¹⁸ alongside the presentation and summary of representative geoscience pre-trained foundation models.

Progress and application of large models in geoscience

The advent of large language models, prominently illustrated by ChatGPT, has significantly advanced diverse domains, concurrently empowering AI technologies to facilitate remarkable scientific progress, notably in geoscience. This is achieved through the autonomous calibration of billions of parameters during training, thereby enhancing representational capacity and learning capability.^{68,419,420,421,422} The application of large models in geoscience, despite its unique challenges, has already demonstrated its huge revolutionary potential over traditional methods,^{18,146,171,211,419,423,424,425} with the most noteworthy advances in the fields of remote sensing, atmosphere, ocean, and hydrology.^{323,426,427,428,429,430}

Specifically, the remote sensing domain owns the most diverse data in the entire Earth science field.^431,432 General applications such as object detection, semantic segmentation, scene classification, and change detection from various data sources promoted the development of large models in remote sensing, such as the largest spectral remote sensing foundation model,⁴³³ with an effective method for expanding and fine-tuning ViT.⁴³⁴ Recently, AI Earth—based on a universal segmentation model (AIE-SEG) —was proposed by Alibaba to quickly extract any target in remote sensing images, achieving unified image segmentation tasks and rapid extraction of “zero samples of all things” without any labeled data. A new AI model called “segment anything model” from Meta AI can “cut out” any object in any image with zero-shot generalization to unfamiliar objects and images, without the need for additional training.⁴³⁵ IBM and NASA have also teamed up to develop an open-source, geospatial foundation model that will enable researchers and scientists to utilize AI to track the amount of satellite data.⁴³⁶ Furthermore, there is rapid development in multimodal remote sensing large models. For instance, SkySense¹⁵⁸ is a generic billion-scale model pre-trained on a curated multi-modal remote sensing imagery dataset with 21.5 million temporal sequences. In addition, large-scale vision-language models, such as EarthGPT,⁴³⁷ have garnered significant attention in the remote sensing field, aiming to unify various remote sensing tasks and multi-sensor images. In a general sense, it can be observed that the utilization of large computer vision models and the efficient exploitation of vast remote sensing datasets to enhance the recognition of various targets represents a prominent trajectory in the evolution of large remote sensing models.

In the climate and weather domains, numerous large models with a great amount of data and parameters have been trained for predictions. For example, a Fourier forecasting neural network (FourCastNet) is proposed to provide immediate accurate short to medium-range global weather predictions.⁴³⁸ The predictive outcomes derived from the FourCastNet model have been meticulously juxtaposed with the findings of the integrated forecasting system. It has been ascertained that the FourCastNet model exhibits substantial advantages across a multitude of performance indicators, with a particular emphasis on its notable progress in the domain of precipitation forecasting. Notably, the accuracy of the FourCastNet model surpasses that of other ones by an impressive margin, exceeding 20%. Pangu-Weather,¹⁸⁷ which harnesses the power of the 3D Earth-specific Transformer, has been empirically demonstrated to yield superior results, accompanied by a remarkable acceleration of 10,000 times, in contrast to the ECMWF. The proposal of NowcastNet,²³² a nonlinear nowcasting model for extreme precipitation, signifies a novel approach that unifies physical-evolution schemes and conditional-learning methods within a neural network framework. This model has proven its capacity to skillfully forecast extreme precipitation events characterized by advective or convective processes, previously deemed challenging to predict. MetNet-3,⁴³⁹ a collaborative development by Google and DeepMind, has enhanced high-resolution predictions of several weather variables, encompassing precipitation, surface temperature, wind speed, and wind direction, for a forecast horizon extending up to 24 h. GenCast⁴⁴⁰ proposes a generative model for global medium-range ensemble weather forecasting up to 15 days ahead, utilizing a diffusion model to sample ensembles from future weather trajectories. In addition, the swift advancement of large language models has positively impacted climate-related endeavors. For example, ClimateGPT⁴⁴¹ serves as a specialized conversational agent for climate change and sustainability topics in English and Arabic.

Concurrently, there have been recent propositions in the development of general geoscientific large-scale models. In the context of hydrology, a foundation platform, HydroPML,³²³ is proposed for hydrological applications based on physics-aware ML. It bridges the gap between large language models and process-based hydrology, offering a range of applications, including but not limited to rainfall-runoff-inundation modeling,¹²² real-time flood forecasting,³²¹ and cutting-edge methods to enhance water security and foster resilient water management. The first-ever large language model in the ocean domain, OceanGPT,⁴²⁹ is introduced as an expert in various ocean science tasks. In the domain of disaster management and response, Disaster Response GPT is proposed to provide a versatile and adaptive framework for addressing various types of disasters and their associated challenges.⁴⁴² Furthermore, large models for time series forecasting, including variables such as wind and weather, have been proposed, leveraging a transformer backbone and zero-shot transfer.⁴⁴³

In summary, substantial advancements have been made in remote sensing and climate domains by deploying large models and effectively utilizing extensive datasets. However, widespread adoption of these methods on a broad scale remains challenging, particularly in extreme weather prediction. Progress in other geoscience areas, such as disaster prevention and hydrology, has been hindered by limited access to datasets and computing resources, slowing down the development of large language models. In the future, developing a unified, interpretable, and continuously learning large model to address the complexities and scales of geoscience will be a focus of ongoing exploration.

Pre-training of large geoscience models

Table 1 illustrates the schematic representation of the foundation of pre-trained models in geoscience. In the realm of remote sensing, various approaches have emerged, for instance, MoCo-V2 with geographic location serving as an agent task in conjunction with contrast learning for base model training,¹⁵⁴ CSPT using knowledge migration and image mask learning to enhance the expressive capability of the pre-trained model,⁴⁴⁴ SeCo constructing positive-negative sample pairs from different seasons to effectively utilize unlabeled multi-seasonal data.⁴⁴⁵ Wuhan University introduced the Billion Visual Transformer model,⁴⁴⁶ exploiting a masking strategy for pre-processing, and achieved notable performance in image classification, target detection, and semantic segmentation. SatMAE,⁴⁴⁷ proposed by Stanford University, adopts a grouped masking strategy for multi-temporal and multi-channel multispectral images. Recently, Hong et al.⁴³³ designed the first and largest customized foundation model for spectral remote sensing data, i.e., SpectralGPT, achieving state-of-the-art performance in various downstream applications. Simultaneously, the work⁴⁴⁸ combines SAR and multispectral images for a contrast learning approach. Another study⁴⁴⁹ employs contrast learning, image filling, and deformation prediction as agent tasks to enhance the generalization of the pre-trained model. Researchers at the University of California, Berkeley focus on spatial scale information, modeling low-frequency and high-frequency details separately in the reconstruction layer.⁴⁵⁰ In addition, Hong et al.⁴⁵¹ explored multimodal fusion on various image types, including optical images, SAR images, digital elevation models, and MAP data,⁴⁵² which innovated a new paradigm of multimodal AI big models for Earth observation, unlocking the Earth observation capability of remote sensing big data.⁴⁵³ Presto reconstructs time series images through stacking and employing randomized masking strategies. Furthermore, GFM employs a teacher-student two-stream network on large-scale datasets,⁴⁵⁴ excelling in scene classification, change detection, and semantic segmentation. Satvit explores the role of the MAE framework in analyzing satellite remote sensing data.⁴⁵⁵

Table 1.

Representative pre-trained foundation models in geoscience

Application field	Model	Pre-trained model	Objectives
Remote sensing	CSPT444	ViT	improving the expressive ability of the pre-trained model
	RingMo456	ViT/Swin Transformer	a remote sensing foundation model with masked image modeling
	Scale-MAE450	Transformer	a pre-trained framework that introduces scale invariance into encoders that are used for a diverse set of downstream tasks
	SatMAE447	Transformer	pre-training Transformers for temporal and multi-spectral satellite imagery
	pre-trained ViT427	ViT	remote sensing foundation model
	GFM454	ViT	building geospatial foundation models via continual pre-training
	SatViT455	ViT	pre-training transformers for Earth observations
	Masked ViT457	ViT	self-supervised masked image reconstruction to advance transformer models for hyperspectral remote sensing imagery
	SpectralGPT433	ViT	the first customized foundation model designed explicitly for spectral remote sensing data
Weather and climate	Earthformer458	Transformer	a space-time Transformer for Earth system forecasting
	FourCastNet438	Fourier Neural Operator	provide accurate short- to medium-range global predictions
	GraphCast157	GNN	medium-range global weather forecasting
	NowcastNet232	physics-conditional generative network	a nonlinear nowcasting model for extreme precipitation
	MetNet439	U-Net + ViT	high-resolution predictions of several core weather variables
	Pangu-weather187	3D Transformer	accurate medium-range global weather forecasting
	ClimateX428	ViT	a foundation model for weather and climate
Others	K2459	Generative model (LLaMA-7B)	Earth science large language model
	DisasterResponseGPT442	Generative model	provide a versatile and adaptive framework for disasters
	OceanGPT429	Generative model	a large language model for ocean science tasks

In a distinct domain, ClimaX is pre-trained on the CMIP6 climate dataset,⁴²⁸ offering versatility in weather and climate tasks. Notably, K2,⁴⁵⁹ a 7 billion parameter Earth science language model from Shanghai Jiao Tong University, utilizes a two-stage construction involving pre-training on a high-quality Earth science corpus and instruction fine-tuning with a geosignal dataset. In contrast, general visual models such as Sky Eye and SenseEarth 3.0 improve remote sensing interpretation efficiency, leveraging Transformer-like backbones and self-supervised learning.

In summary, algorithms designed for processing remote sensing images exhibit variations in their emphasis on RGB, multispectral, or hyperspectral data, tailored for application to specific downstream tasks. Notably, contemporary climate and geoscience models such as K2 and ClimaX exemplify advancements in addressing challenges within these domains, showcasing enhanced efficiency and robustness for applications in Earth science. Despite the immense potential of large geoscience models, common research teams (usually small groups) encounter numerous impediments in embracing large-scale (pre-trained) models. Chiefly, constraints in resources, encompassing limited funding and manpower, impede their capacity to conduct research and development effectively. In addition, the intricacy of large-scale models poses a formidable learning curve for small teams, who may grapple with acquiring expertise across diverse disciplines such as ML and natural language processing. In the long term, the absence of access to comprehensive datasets and formidable competition from large tech companies further impede their progress. Legal and ethical considerations also present challenges, as small teams may lack the resources to adeptly navigate intricate issues such as privacy and accountability. Overall, surmounting these hurdles will necessitate strategic investments, collaboration, and concerted efforts to address legal and ethical concerns.

Deep-time digital Earth

Delving into the deep-time history of Earth is seen as a promising avenue to unravel the mechanisms of Earth’s evolution, expose climate change patterns, identify natural resources, and envisage the future of our planet.^171,460 The advent of big data science in recent decades provides a valuable opportunity to tackle these questions. To expedite exploratory studies of Earth’s evolution, there is a pressing need for an equitable, integrated database. To achieve this goal, the Deep Time Digital Earth (DDE) project is proposed as the inaugural “large-scale scientific project” by the International Union of Geological Sciences. This initiative aims to facilitate deep-time, data-driven discoveries through collaborative efforts across nations and disciplines.⁴⁶¹ Moreover, it introduces an open data platform to establish connections between existing deep-time geocounts and integrated geological data.

• (1)Earth’s life evolution. The synergy of AI and data science has significantly advanced our comprehension of Earth’s life evolution, particularly concerning early complex life and mass extinctions. For instance, ML methods are employed to analyze deep-time marine Paleozoic data, unraveling the impact of environmental changes on biodiversity.⁴⁶² The pulsed extinction of early complex life was further corroborated through network analysis of Ediacaran fossils.⁴⁶³ Furthermore, the DDE project aims to integrate and interconnect existing deep paleontological and stratigraphic databases, leveraging DL and other AI tools to expedite biological data-driven discoveries.⁴⁶⁴
• (2)Earth’s material evolution. In the context of Earth’s material evolution, current AI-driven approaches strive to propel the evolution and discovery of minerals, rocks, sediments, and fluids. Noteworthy examples encompass the evolution of minerals,⁴⁶⁵ the cycling of sediments,⁴⁶⁶ and the interpretation of plate tectonics.⁴⁶⁷ In addition, AI-driven discovery necessitates the integration of existing geomaterial databases by the DDE, enhancing spatial and temporal coverage as well as resolution in the discovery of geomaterials.
• (3)Geography’s evolution. Geography’s evolution holds paramount significance in various domains, including mineral and energy resource assessment, Earth hazard prediction, comprehending Earth’s history, and forecasting the future. The correlation of deep Earth science databases with paleogeographic reconstruction databases is an important goal of DDE. Supported by big data analysis techniques, this combination has been widely used in the field of paleontology,⁴⁶⁸ paleoclimatology,⁴⁶⁹ and geodynamics.⁴⁷⁰
• (4)Paleoclimate’s evolution. The exploration of paleoclimate assumes a crucial role in understanding the interaction between Earth and life in producing climate extremes and forecasting future climate changes.⁴⁷¹ AI’s strengths in data processing, hypothesizing, and predicting within Earth science research substantially facilitate paleoclimate reconstruction.⁴⁷² Assisted by AI, the DDE can reconstruct the history of paleoclimate and paleoatmosphere, relying on various minerals, rocks, and geochemical indicators preserved in Earth material.⁴⁷³

In summary, the establishment of a unified representation model to head the construction of an integrated Earth science knowledge map is one of the key programs of DDE,^474,475 and a series of knowledge graphs have emerged, such as the paleoclimate knowledge graph,⁴⁷⁶ standard carbonate microfacies,⁴⁷⁷ and academic knowledge graph.⁴⁷⁸ With the continued emergence of geoscientific macrolanguage models (such as K2⁴⁵⁹), AI has dramatically changed the traditional paradigm of geoscientific research. By harmonizing and integrating deep Earth data, geological knowledge, and advanced techniques in data science and AI, DDE is poised to advance solutions for the significant challenges in Earth evolution research, understanding the past, present, and future of our planet.

Challenges and outlooks in AI for geoscience

The numerous cases and advanced techniques outlined in the previous sections solidly prove that AI is an expert technology at deciphering complex relationships in the Earth system and predicting environmental responses with unprecedented accuracy. However, this is not the end of the journey; there remain ongoing challenges and opportunities in the field of research. This section poses the challenges and future perspectives to promote the co-development of AI and geoscience.

Unsolved challenges of AI for geoscience

There are many unsolved challenges in AI for geoscience, particularly at the intersection of these two fields. These challenges arise from interdisciplinary complexities, making it difficult for scientists to identify and address the problems.

Ethical considerations play a crucial role across all stages of geoscience disciplines, encompassing data collection, analysis, and distribution.⁴⁷⁹ High-resolution data, for example, raise privacy concerns,⁴⁸⁰ while socio-economic analyses can lead to stigmatization if not handled carefully.⁴⁸¹ The demand for explainability grows as AI applications extend their reach into policy-making, requiring models to be both transparent and justifiable.¹⁴⁶ Addressing these ethical challenges involves adhering to robust ethical frameworks and guidelines, promoting a culture of geoethical thinking and social responsibility among researchers.

Moreover, due to the biased learning knowledge by AI, the adeptness of AI in modeling complex relationships brings about vulnerabilities related to data security.^482,483 The potential for data bias and tampering poses significant risks, potentially leading to misrepresentations of geographical features and misguided policy decisions. To mitigate these risks, a multifaceted approach, including robust data validation and enhancements in AI learning specifications, is essential. These strategies not only fortify data integrity but also improve the resilience of AI systems against malicious manipulations.

Despite the exceptional capabilities of AI, the demand for computing resources and the costs associated with data acquisition and processing present substantial challenges.¹⁷ The computational intensity required for models, such as predicting global climate¹⁸⁷ or global forest fire interactions,⁴⁸⁴ necessitates substantial investment in computational and memory resources, often beyond the reach of many geoscientists. Moreover, the AI models should be energy efficient so that they can also contribute to the NetZero agenda. To optimize performance and reduce expenses, strategies such as leveraging cloud computing, applying transfer learning, and enhancing data management practices are vital.²⁸⁴ These approaches help in managing the high costs and logistical demands of extensive data processing, ensuring that AI applications remain both viable and effective.

Emerging challenges in new paradigm of hybrid models

Hybrid models, leveraging the strength of physics-based models and AI, are starting to show their charming potential as a new research paradigm in geoscience. Despite their potential, they present challenges in the development of the paradigm.

The first challenge is the uncertain interpretability within the model. While the structure of hybrid models seems to maintain physical plausibility, and the AI component can even effectively compensate for structural deficiency in physics-based counterpart,^485,486 there remains a critical concern. Often, the balance between physics-based and AI components in hybrid models may be overlooked due to a lack of integration knowledge within the “gray box.” The work by Acuña Espinoza et al.⁴⁸⁷ suggests that AI-based parameterization may learn incorrect behaviors and overwrite the physical interpretability in the hybrid hydrological models, despite enhancing performance. This compensatory capability of AI raises questions about the true hydrological interpretability of outputs from hybrid models. It also calls for a more cautious use of hybrid models in geoscience applications, particularly when the primary objective is to decipher geophysical processes rather than merely improve prediction accuracy.

Another challenge in advancing this paradigm is extending these hybrid models to accommodate large datasets and complex system interactions inherent in global geoscience applications. As these models scale, the structural deficiencies in the physics-based part of the hybrid model will be magnified,⁴⁸⁵ and maintaining a balance between AI fitting capabilities and physical interpretability will become increasingly difficult. Therefore, large models currently applied in geoscience, such as the FourCastNet and Pangu-Weather models, are still predominantly in the data-driven paradigm and risk losing physical plausibility. This scaling issue highlights the need for a deep understanding of geophysical processes in hybrid models at the regional scale.

Outlook on AI for inter-spheres

While the application and knowledge of AI for intra-spheres are relatively comprehensive, exploring inter-spheres connection in geoscience reveals significant knowledge gaps.⁴⁸⁸ These gaps arise from the challenges of integrating fragmented knowledge across disciplines when enhancing Earth system models. The complexity of cross-system dynamics and feedback mechanisms complicates the encoding of multidisciplinary and multi-domain knowledge. For instance, the biochemical and biophysical processes within the hydrological cycle⁴⁸⁹ and the atmospheric-ocean interaction⁴⁹⁰ are crucial cases for understanding the hydrological cycle and predicting phenomena such as the Madden-Julian Oscillation and El Niño Southern Oscillation, respectively. Yet, they exhibit gaps in multidisciplinary integration.

Undoubtedly, AI has demonstrated the potential to bridge these interdisciplinary gaps, as demonstrated by its successful application within individual domains. Several studies have already started to apply AI to forge connections across multiple spheres. For example, AI-powered prediction models have been used to forecast hurricanes by analyzing the complex interplay between ocean temperatures, atmospheric conditions, and land surface characteristics.⁴⁹¹ However, advancing AI development in the inter-sphere’s context requires greater efforts, including more robust exchanges of expert knowledge and domain-specific insights.

Outlook of AI for exploring exoplanets

The lack of terrestrial data with viable and varied observational evidence represents a significant bottleneck in the development of geoscience. Terrestrial exoplanets, sharing similar geophysical processes, can complement the data gap. Planetary scientists suggest that the understanding of the cooling and transfer of heat from the interiors of terrestrial planets can help explain the geological evolution of Earth.⁴⁹² Furthermore, studying tidal interaction on low-mass planets can aid in understanding atmospheric circulation and meteorological phenomena on Earth.⁴⁹³ This highlights the potential of exoplanet exploration to offer new insights into our own planet.

In contrast to knowledge transfer from exoplanets to Earth, there remain plenty of unknowns about the environment of exoplanets, frequently resulting in a less sophisticated understanding of their geophysical processes compared with Earth. Generally, discussions about exoplanet characteristics often simply rely on the knowledge of an exoplanet’s mass, radius, or orbital distance. In this context, the power of AI can be used to decipher the high complexity of an exoplanet’s system. Some works^494,495 suggest that AI approaches trained by biosignatures on Earth could be adapted for searching for life on terrestrial exoplanets. Interdisciplinary application of AI in geoscience, transferring from Earth to exoplanets, could enhance our understanding of these distant worlds’ geophysical processes, thereby offering a fresh perspective on Earth in the future.

Future development of AI for geoscience

Our review demonstrates the necessity of advancing AI for geoscience research. Looking ahead, AI is poised to significantly enhance geoscience projects, supported by various government and authoritative endorsements. For example, the China Ministry of Science and Technology highlights AI as a pivotal tool for groundbreaking research across four strategic frontiers: deep space, deep sea, deep Earth, and “deep blue.” Similarly, NASA regards AI as an essential component for future Earth explorations.⁴⁹⁶

Conversely, our review also acknowledges the profound and dynamic impact of AI on our understanding of geoscience and on decision-making processes. However, there is limited consensus on the regulations governing AI development and usage. The United Nations Educational, Scientific and Cultural Organization⁴⁹⁷ and the European Union’s General Data Protection Regulation⁴⁹⁸ underscore the importance of ethical considerations, such as privacy, interpretability, and security in AI applications, which indicates the need for a model-data-driven paradigm to enhance transparency in research.

Conclusion

The research paradigms in geoscience started with physics-based models, followed by data-driven approaches, and merged into hybrid models. This review strives to delineate these paradigms, emphasizing the unexplored frontiers where cutting-edge AI techniques intersect with geoscience. We put a special focus on hybrid models, which, leveraging domain knowledge to guide AI models, often require less training data while maintaining comparable accuracy, thus offering enhanced efficiency and performance. The potential of large-scale AI models in geoscience is vast, yet its realization faces challenges unique to the domain, impeding its widespread adoption and implementation. The dichotomy between these paradigms—space centered on explicit adherence to physical rules versus the extraction of insights from immense data volumes—underscores the need for a balanced approach in contemporary geoscience.

In essence, the quest to comprehend Earth’s intricacies demands an amalgamation of diverse methodologies and approaches. The synergy between traditional principles and modern AI-driven techniques holds immense promise, yet it also presents a spectrum of challenges that require concerted efforts to overcome. As geoscientists navigate this dynamic terrain, a harmonized blend of methodologies stands poised to unlock profound insights into our planet’s mysteries, shaping the trajectory of geoscience in the years to come.

Acknowledgments

This work was partially supported by National Natural Science Foundation of China (T2225019, 41925007, 62372470, U21A2013, 42201415, 42022054, 42241109, 42077156, 52121006, 42090014, and 42325107), the National Key R&D Programme of China (2022YFF0 500), the Youth Innovation Promotion Association CAS (2023112), the Strategic Priority Research Program of CAS (XDA23090303), and the RECLAIM Network Plus (EP/W034034/1).

Author contributions

C.L., Y. Xie, and A.P. wrote the introduction. M.C., F.Z., and Z.Q. wrote the paradigms section. S.W., L.Y., C.Y., W.H., T.S., Z.S., T.Q., and Z.C. wrote the AI-driven geoscience paradigms section. C.S., S.Y., N.L., and Y.Z. wrote the atmosphere section. H.Z. wrote the lithosphere section. J. Zeng, H.S., C.Z., and J. Zhang wrote the hydrosphere section. T.Z. wrote the cryosphere section. L. Wang, N.H., and C.H. wrote the biosphere section. L.L., H.Z., and W.Z. wrote the other domains section. T.Z., H.L., J.S., and D.F. revised the typical cases section. C.O., Q.X., Y.W., S.W., and D.H. wrote the large models in geoscience section. J. Zhang, Z.W., Y.L., and T.Z. wrote the challenges and future perspectives in AI for geoscience section. A.P., Lizhe Wang, Y. Xu, F. Wang, B.Z., P.K., and J.L. revised the paper.

Declaration of interests

The authors declare no competing interests.

Published Online: August 22, 2024

Lead contact website

https://grzy.cug.edu.cn/LizheWang.