On August 17, 2021, the journal Nature Machine Intelligence published an article titled "Towards neural earth system modelling by integrating artificial intelligence in earth system." science）。 Based on the systematic analysis of Earth system models (ESMs) and artificial intelligence (AI), the Geoscience Center for Earth Sciences (GFZ) in Helmholz, Germany, has creatively proposed a new concept of "neural earth system modelling". On September 8, 2021, the American Association for the Advancement of Science (AAAS) published a review article titled "Opportunities and limits of AI in climate modelling" to focus on the results.
1 Classical Earth system modeling and its limitations
Earth system models (ESMs) are the most important tool for quantitatively describing geophysical states, such as climate models, which predict how the Earth will change in the future under the influence of human activities. Classical models of the Earth system are based on the well-known laws of physics, and with the help of mathematical and numerical methods, the state of the system in the future time is calculated based on the current or past known state of the system. Despite the great success of ESM, long-standing problems and uncertainties remain: (1) A key measure of evaluating ESMs is equilibrium climate sensitivity, defined as the balance increase in global average temperature due to the transient doubling of atmospheric CO2. The current ESM still has a large range of equilibrium climate sensitivity. From CMIP5 to CMIP6, the range of possible equilibrium climate sensitivity expands from 2.1 to 4.7 °C to 1.8 to 5.6 °C. Reducing these uncertainties, and thus uncertainties in future climate projections, is one of the key challenges in developing ESM. (2) Both theoretical studies and paleoclimate data suggest that several subsystems of the Earth system can abruptly change their state in response to a gradual change in forcing. There are concerns that current ESMs will not be able to predict sudden future climate change because less than 2 centuries of instrument age have not undergone a similar shift, and because of the length of the relevant timescale, model validation of such events using paleoclimate data is still impossible. In extensive studies, many relatively abrupt events have been identified in future projections of cmIP5 models, but due to the nature of these rare, high-risk events, the accuracy of ESMs predicting them remains to be tested. (3) Current environmental stabilization mechanisms are not yet suitable for assessing the effectiveness or environmental impact of CO2 removal technologies, which are considered key to achieving the goals of the Paris Agreement. In addition, ESMs do not adequately reflect key environmental processes such as carbon cycling, the effectiveness of water and nutrients, or the interaction between land use and climate. This could affect the effectiveness of land-based mitigation options that rely on biomass energy for carbon capture and sequestration or actions such as nature-based climate solutions. (4) The dynamic distribution of the Earth system encoded by time series has a typical heavy tail feature. Extreme events, such as heat waves and droughts, as well as extreme precipitation events and associated flooding, always cause huge socio-economic losses. Such events are expected to become more severe as anthropogenic climate change continues, and extreme event attribution will pose another prominent challenge to Earth system science. While current ESMs are very skilled at predicting averages of climate volumes, there is still room for improvement in representing extremes.
2 Advances in artificial intelligence
The challenges of classical ESM methods, and the increasing availability of Earth observations, open up areas for the use of AI. This includes machine learning (ML) methods such as neural networks, random forests, or support vector machines (SVMs). The advantage is that they are self-learning systems that do not require an understanding of physical laws and relationships that may be very complex or even unknown. Instead, they are trained on large data sets for specific tasks and learn about the underlying system itself. This flexible and powerful concept can be extended to almost any required complexity. For example, neural networks can be trained to recognize and classify patterns in satellite imagery, such as cloud structures, ocean eddies, or crop quality. Or learn to weather forecast based on previous records, models, and equations of physical equilibrium. However, to what extent this self-learning approach can truly extend or even replace classical modeling methods remains to be seen. Because machine learning also has its pitfalls, there are many proof-of-concept studies in machine learning applications in climate science today that work in simplified environments. Further studies will judge its operability and reliability. Another decisive aspect, as in the black box, is that the inputs and outputs are well known, but the process of acquiring knowledge behind them is not well known. This can lead to problems with the physical consistency of the validation results, even if they seem to make sense. Interpretability and explainability are important issues in the context of machine learning, and these need to be improved in the future to enhance transparency and trust in methods.
3 Neural Earth System Modeling
A team of mathematicians fused the above two methods into "neural Earth system modeling." Machine learning is no longer used solely for pure data analysis, but also for taking over or speeding up certain process steps within the framework of classic ESM. In this way, the computing power is freed up, further improving the model. For ease of comparison, the researchers distinguished between weakly coupled NESYM hybrid models (where ESM or AI techniques benefit from their respective information) and strongly coupled NESYM hybrid models (fully coupled model-network combinations that dynamically exchange information). The weakly coupled NESYM hybrid model is primarily designed to address the ESM limitations described earlier, particularly unsolved and sub-mesh-scale processes. In the strongly coupled NESYM hybrid model, where models and AI tasks are transposed, the flow of information is directly from the model to the AI tool. Here, neural networks are trained directly with model state variables, their trajectories, or more abstract information such as seasonal signals, interannual cycles, or coupling mechanisms. The goal of a machine learning application may be not only model simulation, but also inversion of nonlinear geophysical processes, learning geophysical causation or predicting extreme events. The evolving process of method fusion will allow a mixture of neural networks, ESMs, and ESOs to exchange information dynamically. ESM will soon use output from supervised and unsupervised neural networks to optimize their physical consistency, which in turn will feed back improved information back to ml components. The researchers developed the key characteristics of NESYM and defined goals: (1) it was possible to replicate and predict out-of-distribution samples and extreme events; (2) perform constrained and consistent simulations, complying with physical protection laws despite potential flaws in individual components; (3) mixing comprehensive adaptive measures that included self-validation and self-correction; and (4) allowing for reproducibility and interpretability
The researchers look forward to the novel interfaces that could establish a dynamic exchange of information between the two methods, thereby continuously improving each other. This deep expansion of process-based classical Earth and climate research elevates neural Earth system modeling to a new, rapidly emerging branch of research. At its core is a hybrid system that can test, correct, and improve its physical consistency, enabling more accurate prediction of geophysical and climate-related processes. Still, AI and hybrid approaches carry high risks and pitfalls, and it's far from clear whether the hype surrounding AI use will, at least on its own, solve the problem of openness. However, in any case, it is worth taking this road. However, to achieve this, it will become increasingly important to have climate and earth research on the one hand and close cooperation between AI experts on the other.
Please indicate the source and author when reprinting this article: Earth Science Dynamic Monitoring Express, Lanzhou Documentation and Information Center, Chinese Academy of Sciences, No. 18, 2021, compiled by Liu Wenhao.