AI Trends

For AI to Understand the Earth, Data Alone Is Not Enough

Sina TechnologyUpdated Nov 19 09:36

AI summary

Experts at a Chinese Academy of Sciences forum noted that AI and big data can help Earth system science advance from observation to prediction and simulation, but cannot replace scientific mechanisms, causal testing, or value judgments. Research driven jointly by data and physical mechanisms has extended early warnings for flash floods and debris flows to around two hours in advance, while humans should also be incorporated into the Earth system as an endogenous variable.

Body · AI translation

Add new public beta promotional copy Add new public beta promotional copyend Add app-exclusive traffic diversion

■By Sun Tao, Staff Reporter

The 243rd Science and Technology Frontier Forum of the Academic Divisions of the Chinese Academy of Sciences, themed “Empowering Innovation in Earth System Science Through Digital Intelligence,” was recently held in Zhengzhou, Henan Province. Academicians and experts from fields including Earth science and information science discussed theoretical innovation in Earth system science and the application of digital-intelligence technologies in human–Earth systems, climate, oceans, the deep Earth and other areas.

From Seeing the Earth Clearly to Simulating Its Future

“The pursuit of digitally driven Earth system science involves four steps: first seeing the Earth clearly, then understanding it, then predicting it and finally simulating its future,” said Guo Huadong, an academician of the Chinese Academy of Sciences, outlining the direction of digitally empowered Earth system science through four progressive goals.

From data to information, from information to knowledge and then to intelligence, “Earth intelligence” has become the new main thread of global Earth observation research. The focus of Earth science is shifting away from traditional studies of a single sphere or local process and toward the complex interactions within the Earth’s natural systems and between those systems and human activities. Advances in big data and AI have enabled scientists to acquire and process previously unimaginable volumes of data, explore intelligent approaches to understanding the Earth, and provide fresh impetus for understanding individual spheres, multispherical coupling and the Earth system as a whole.

Guo believes that AI and big data cannot be regarded merely as two auxiliary tools external to Earth science. “Big data and AI empower each other, and their integration with the Earth system creates Earth system science intelligence. This will transform the research paradigm of Earth system science and enable humanity to understand planet Earth more deeply and systematically.” However, growth in the scale of data does not mean that scientific understanding will automatically deepen.

Citing the Alpha series of achievements in AI, Li Xin, a researcher at the Institute of Tibetan Plateau Research of the Chinese Academy of Sciences, posed a question: When will Earth system science have its “Alpha moment”?

AlphaGo can find optimal strategies in the clearly defined world of Go, and AlphaFold can predict protein structures, but is there a complete mathematical theory capable of supporting simulation and decision-making for the complex Earth system? Li believes that the Earth system is characterized by multiple scales, strong nonlinearity, heterogeneity among its spheres and openness. Its rules are neither fully transparent, as they are in Go, nor can all possibilities be exhausted through limited training data.

He noted that if the core of the traditional paradigm lies in “harmony between dynamics and observation,” then the key to empowering Earth governance with digital-intelligence technologies is to let “Kalman meet Bellman”—integrating state estimation, or data assimilation, represented by Kalman filtering with long-term value-based decision optimization, or reinforcement learning, represented by the Bellman equation.

Writing Humanity Into the Earth System

Earth system science must study not only nature but also humanity.

Fan Jie, an academician of the Chinese Academy of Sciences, noted that traditional human geography relies mainly on statistical induction. When confronted with complex human behavior and social relationships, it is difficult to model them fully using traditional numerical equations. Today, combining information such as policy documents, socioeconomic data and population movement trajectories with AI technologies offers new possibilities for studying human–Earth systems.

Technological advances, however, have also created new problems.

Human behavior is not a simple form of physical motion. Different groups have different interests, values and behavioral choices. A model may be able to calculate the efficiency of a particular plan, but it may not be able to determine whether that plan meets the actual needs of social development.

Fan warned that values often play a decisive role in the process of sustainable development. If foundation models lack guidance from a scientific worldview and sound values, or if they overlook the complexity and diversity of social equity, algorithmic simulation alone will struggle to truly solve the central challenges of sustainable development in human–Earth systems.

Liu Yansui, a professor at Henan University, proposed a theoretical framework for the modern “human–Earth sphere,” explaining a frontier direction for innovative development in human–Earth system science. In his view, traditional Earth system science often treats humanity as an outside observer or external influencing factor. Genuine research into the Earth’s surface system, however, requires an epistemological shift: humanity must change from being an observer and influencer of the Earth system into an endogenous variable and core component of the system.

He also proposed the concept of “geographical engineering”: If medicine treats diseases of the human body through “taking medicine, receiving injections and undergoing surgery,” then geographical engineering addresses real-world problems such as degraded human–Earth systems, vacant and abandoned rural land, desertification and salinization through “conservation, restoration and engineering.” This approach extends Earth system research beyond understanding and protecting nature into the new dimensions of transforming nature and coupling human and Earth systems. Liu emphasized that modern geographical research cannot stop at writing papers at a desk or building models indoors; researchers must “conduct scientific research in the fields and write their papers across the land of our motherland.”

The value of digital-intelligence technologies must ultimately be reflected in their ability to solve practical problems.

Cui Peng, an academician of the Chinese Academy of Sciences, discussed research on monitoring and providing early warnings of flash-flood and debris-flow risks, driven jointly by data and physical mechanisms.

In the past, disaster warnings relied more heavily on meteorological rainfall forecasts. But rainfall is not equivalent to disaster risk: the same amount of rain may have entirely different consequences under different terrain, soil-moisture and geological conditions.

Cui’s team connected processes including rainfall, soil-moisture infiltration, soil instability, debris-flow initiation and downstream movement. By combining them with spatiotemporal big data for dynamic risk assessment, the team explored refined forecasting and early warning for mountain hazards, extending the advance warning time for flash floods and debris flows to around two hours while substantially improving forecast and warning accuracy.

Cui Peng noted that traditional forecasts and early warnings often focus on whether a disaster will occur, but what disaster prevention and mitigation really need to know is not only whether a disaster will occur, but also what harm it may cause. This requires linking rainfall forecasts with disaster formation and the resulting losses. In addition to accurate rainfall forecasts, it is necessary to consider how water and soil interact after rain reaches the ground, when the soil will become unstable and trigger landslides, and where debris flows will evolve into disasters, thereby precisely identifying hazardous areas and risks to exposed elements and enabling early warnings and evacuations.

In climate change research, Zhou Tianjun, a researcher at the Institute of Atmospheric Physics of the Chinese Academy of Sciences, briefly reviewed the international development of climate system models, with a focus on the super-ensemble simulation method, and discussed how AI could advance climate system simulation research in the future.

Zhou Tianjun emphasized that climate observations contain both changes caused by human activity and the climate system’s inherent natural variability. Through a large number of simulation experiments, super ensembles can help scientists distinguish the “signal” of human influence from the “noise” of natural fluctuations and analyze the physical mechanisms behind anthropogenic changes and natural oscillations, providing a basis for climate change detection, attribution and projection.

Ocean research is undergoing similar changes. Li Xiaofeng, a researcher at the Institute of Oceanology of the Chinese Academy of Sciences, presented research progress on the “Langya” large ocean model. In the past, ocean forecasting focused mainly on state variables such as temperature, salinity and ocean currents, while new research is expanding into complex ocean phenomena including typhoons, precipitation, storm surges, internal solitary waves, mesoscale eddies and sea ice.

In deep-Earth research, Zheng Yongfei, an academician of the Chinese Academy of Sciences, emphasized the importance of scientific mechanisms. “Qualitative understanding comes first; quantification comes second.” He noted that if the essential nature of a geoscientific problem is misjudged, even the most complex models and powerful computing resources may produce misleading results.

As Guo Lei, an academician of the Chinese Academy of Sciences, emphasized during the forum’s closing discussion, a rooster’s crow and sunrise are highly correlated in observations, but the crowing cannot be regarded as the cause of sunrise. Similarly, algorithms can help scientists discover complex associations in vast amounts of data and can even assist in generating causal hypotheses and designing experiments; but penetrating beyond correlation to establish causation usually also requires domain theory, causal identification strategies and validation using multiple lines of evidence. Statement: This is an exclusive Sina.com article and may not be reproduced without authorization. tech_web_article_v2015_block_bottom Non-targeted 300*250 button 17/09 wenjing begin

Non-targeted 300*250 button end

Source

Sina Technology