STC: Center for Learning the Earth with Artificial Intelligence and Physics (LEAP)

STC：利用人工智能和物理学习地球中心 (LEAP)

• 批准号: 2019625
• 负责人: Pierre Gentine
• 金额: $ 2500万
• 依托单位: Columbia University
• 依托单位国家: 美国
• 项目类别: Cooperative Agreement
• 财政年份: 2021
• 资助国家: 美国
• 起止时间: 2021-10-01 至 2026-09-30
• 项目状态: 未结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=2019625&HistoricalAwards=false
• 关键词: STC; Center; Learning; Earth; Artificial

Projections of future climate change from Earth system models (ESMs) play a critical role in addressing the threats posed by climate change, especially given the need to plan for conditions that have no historical precedent. But ESM projections have large uncertainties which limit their usefulness for decision support, and the most worrisome forms of climate change are often the ones with the greatest uncertainties. Such uncertainty is not unexpected considering the many processes, from cloud formation to carbon cycling to ocean turbulence, that affect the climatic response to anthropogenic forcing. These processes must be represented in ESMs but there is no easy way to simulate them. Some, like ocean turbulence, are hard simply because they involve spatial scales too small to be simulated at the resolutions accessible to global models. Others, like the exchange of water and carbon dioxide through a forest canopy, are crudely simulated due to incomplete scientific understanding.Processes which cannot be explicitly represented in ESMs are instead incorporated through “parameterizations”, which roughly express their effects on the resolved model state. Such parameterizations are based on first-principles theory but they also involve crude approximations and must be “tuned” by assigning numerical values to parameters which control their behavior. The parameters typically lack observational and theoretical constraints, and their values are manually adjusted to improve the simulation of present-day climate. Even after tuning models still exhibit substantial bias, and the complexity and computational expense of ESMs has increased to the point where traditional hands-on tuning is becoming impractical.Artificial Intelligence (AI), particularly in the form of Machine Learning (ML), offers a new way forward for improving parameterizations and reducing uncertainty in climate projections. AI is a compelling complement to traditional parameterization development, which begins with theory and physical principles and uses observational data somewhat sparingly. In contrast, the methods of AI are data driven and thus a perfect match for the explosive growth in earth system data that has occurred in recent years. This includes data from satellites, in situ networks, and field campaigns. For some processes, particularly cloud formation and ocean turbulence, small-scale process models have become sufficiently realistic that they can provide surrogate observations to drive AI-based methods.The Center for Learning the Earth with Artificial Intelligence and Physics (LEAP) applies the power of AI to the wealth of available earth system data to overcome the limitations of traditional parameterization development and tuning, thus creating a new pathway to better ESMs and better guidance to decision-makers. The AI methods are novel in that they build physical constraints such as conservation laws into data-driven algorithms. AI methods are also used to find more discriminating ways to use observational data to evaluate model performance. LEAP works with the developers of the Community Earth System Model to ensure that its advances are made available to the worldwide community of climate researchers.In order to promote effective climate adaptation, LEAP fosters equitable training of the next generation of diverse learners across multiple scales by supporting post docs, graduate students, high-school students, parents, and teachers. Furthermore, the Center supports bidirectional knowledge transfer with the public and private sectors to develop tailored and relevant climate-related information for stakeholders so that they can better adapt to climate change.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.

地球系统模型（ESM）对未来气候变化的预测在应对气候变化带来的威胁方面发挥着至关重要的作用，特别是考虑到需要为历史上没有先例的条件进行规划。 但ESM预测具有很大的不确定性，这限制了它们对决策支持的有用性，而最令人担忧的气候变化形式往往是不确定性最大的形式。 考虑到从云的形成到碳循环再到海洋湍流等影响气候对人为强迫的反应的许多过程，这种不确定性并不意外。 这些过程必须在ESM中表示，但没有简单的方法来模拟它们。有些，比如海洋湍流，很难，仅仅是因为它们涉及的空间尺度太小，无法以全球模型的分辨率进行模拟。由于科学认识的不完整，其他的过程，如水和二氧化碳在林冠中的交换，只能粗略地模拟，而不能在ESM中明确表示的过程则通过“参数化”来表示，这些参数化可以粗略地表示它们对解析模型状态的影响。 这种参数化是基于第一原理理论，但它们也涉及粗糙的近似，必须通过分配数值的参数，控制他们的行为进行“调整”。 这些参数通常缺乏观测和理论上的限制，它们的值是人工调整的，以改善对当今气候的模拟。 即使在调整之后，模型仍然存在很大的偏差，并且ESM的复杂性和计算费用已经增加到传统的手动调整变得不切实际的程度。人工智能（AI），特别是机器学习（ML）的形式，为改进参数化和减少气候预测的不确定性提供了一种新的方法。 人工智能是传统参数化开发的一个引人注目的补充，传统参数化开发从理论和物理原理开始，并在一定程度上节省使用观测数据。 相比之下，人工智能的方法是数据驱动的，因此非常适合近年来地球系统数据的爆炸性增长。 这包括来自卫星、现场网络和实地活动的数据。 对于某些过程，特别是云的形成和海洋湍流，小尺度过程模型已经变得足够逼真，它们可以提供替代观测来驱动基于人工智能的方法。人工智能和物理学学习地球中心（LEAP）将人工智能的力量应用于丰富的可用地球系统数据，以克服传统参数化开发和调整的局限性，从而为更好的紧急保障措施和更好地指导决策者开辟了一条新的途径。人工智能方法是新颖的，因为它们将守恒定律等物理约束构建到数据驱动的算法中。 人工智能方法也被用来寻找更有区别的方法来使用观测数据来评估模型性能。LEAP与社区地球系统模型的开发者合作，确保其进展成果能够提供给全球气候研究者社区。为了促进有效的气候适应，LEAP通过支持博士后、研究生、高中生、家长和教师，在多个尺度上促进对下一代多样化学习者的公平培训。此外，该中心还支持与公共和私营部门的双向知识转移，为利益相关者开发量身定制的相关气候相关信息，以便他们能够更好地适应气候变化。该奖项反映了NSF的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

项目成果

Development of a Deep Learning Emulator for a Distributed Groundwater–Surface Water Model: ParFlow-ML

• DOI: 10.3390/w13233393
• 发表时间: 2021-12
• 期刊: Water
• 影响因子: 3.4
• 作者: Hoang Tran;E. Leonarduzzi;Luis De la Fuente;R. B. Hull;Vineet Bansal;Calla Chennault;P. Gentine;Peter Melchior;L. Condon;R. Maxwell

Deep Learning for Subgrid‐Scale Turbulence Modeling in Large‐Eddy Simulations of the Convective Atmospheric Boundary Layer

• DOI: 10.1029/2021ms002847
• 发表时间: 2022-05
• 期刊: Journal of Advances in Modeling Earth Systems
• 影响因子: 6.8
• 作者: Yu Cheng;M. Giometto;Pit Kauffmann;Ling Lin;Chengyu Cao;Cody Zupnick;Harold Li;Qi Li;Y. Huang;R. Abernathey;P. Gentine

Variability in the Global Ocean Carbon Sink From 1959 to 2020 by Correcting Models With Observations

• DOI: 10.1029/2022gl098632
• 发表时间: 2022-07-28
• 期刊: GEOPHYSICAL RESEARCH LETTERS
• 影响因子: 5.2
• 作者: Bennington, Val;Gloege, Lucas;McKinley, Galen A.
• 通讯作者: McKinley, Galen A.

Correcting Systematic and State‐Dependent Errors in the NOAA FV3‐GFS Using Neural Networks

使用神经网络纠正 NOAA FV3–GFS 中的系统性和状态相关错误

• DOI: 10.1029/2022ms003309
• 发表时间: 2022
• 期刊: Journal of Advances in Modeling Earth Systems
• 影响因子: 6.8
• 作者: Chen, Tse‐Chun;Penny, Stephen G.;Whitaker, Jeffrey S.;Frolov, Sergey;Pincus, Robert;Tulich, Stefan
• 通讯作者: Tulich, Stefan

Explicit Physical Knowledge in Machine Learning for Ocean Carbon Flux Reconstruction: The pCO 2 ‐Residual Method

海洋碳通量重建机器学习中的显式物理知识：pCO 2 – 残差法

• DOI: 10.1029/2021ms002960
• 发表时间: 2022
• 期刊: Journal of Advances in Modeling Earth Systems
• 影响因子: 6.8
• 作者: Bennington, Val;Galjanic, Tomislav;McKinley, Galen A.
• 通讯作者: McKinley, Galen A.