EAR-Climate: Towards Better Understanding of Global Low Flow Dynamics Under Climate Change With Next-Generation, Differentiable Global Hydrologic Models

淡水资源对世界各地许多相互竞争的人类和生态系统需求至关重要。需要全球水文模型来评估气候变化对水资源的影响，但其准确性往往有限。模型方程有时是任意的，我们没有充分利用现有数据的价值。特别是，传统的模型很难描述河流干涸时的枯水期，有时预测的趋势与观测记录相反。这些错误可能导致气候缓解战略不足、干旱准备不足或救灾资源分配不当。机器学习模型往往是准确的，但它们仍然对人类破译具有挑战性，不太适合提出精确的问题，也不一定尊重我们知道是正确的物理定律，如质量守恒。这项工作将寻求建立一种新的水文模型流派，目前被称为水文学中的可微建模，或简称为可微水文学。从该项目开发的新一代全球水文模型不仅将提高我们估计未来枯水流量的能力，而且这项工作还将建立一条将机器学习和过程的最佳方面结合在一起的水文学新途径。新的途径将提供灵活的方式来提出新的科学问题并从大数据中学习答案。因此，水文学家将不再受到人工智能中通才模型设计的限制，在这种设计中，可解释性被用来换取模型的通用性。为了实现项目目标，将剥离人工智能的各层，以利用其核心技术之一--可区分编程--来构建可学习的基于过程的模型。下一代水文模型将基于全球水文数据演变，减少结构性缺陷，建立全球地下水参数方案，并解决规模问题。该项目将表征可归因于结构缺陷的错误，提高数据稀疏地区预测的可靠性，并提高模型的物理意义。结果将分发给气候变化影响评估界。还将通过与在非洲有足迹的非营利组织继续合作，分享低流量预测。这项研究工作将通过他们的老师被纳入到研究生、研讨会和研讨会与会者以及高中生的教育活动中。该项目由水文科学计划资助，以及地球科学理事会和高级网络基础设施办公室之间的合作，以支持地球科学中的AI/ML和开放科学活动。该奖项反映了NSF的法定使命，并通过使用基金会的智力价值和更广泛的影响审查标准进行评估，被认为值得支持。

Freshwater resources are critically important to many competing human and ecosystem needs around the world. Global hydrologic models are needed to assess climate change impacts on water resources, but their accuracy is often limited. Model equations are sometimes arbitrary and we do not fully leverage the value of available data. In particular, traditional models have trouble describing low flow periods when rivers run dry, and sometimes predict trends that are opposite to observational records. These errors could lead to inadequate climate mitigation strategies, under-preparation for drought, or misallocation of disaster relief resources. Machine learning models tend to be accurate, but they remain challenging for humans to decipher, are not well suited to ask precise questions, and do not necessarily respect physical laws we know to be true such as the conservation of mass. This work will seek to build a new genre of hydrologic modeling currently termed differentiable modeling in hydrology¸ or, simply, differentiable hydrology. Not only will the next-generation global hydrologic models developed from this project improve our ability to estimate future low flows, but this work will also establish a new avenue in hydrology that combines the best aspects of machine learning and processes. The new avenue will provide the flexibly to ask new scientific questions and learn the answers from big data. As a result, hydrologists will no longer be limited by the generalist model design in artificial intelligence where interpretability is traded for model genericity. To achieve the project goals, the layers of artificial intelligence will be peeled off to harness one of its core technologies, namely, differentiable programming, to build learnable process-based models. Next-generation hydrologic models will evolve based on global hydrologic data, reduce structural deficiencies, build global parameterization schemes for groundwater, and address scale issues. The project will characterize errors attributable to structural deficiencies, improve reliability of predictions in data-sparse regions, and improve model physical significance. Outcomes will be disseminated to the climate change impact assessment community. Low flow predictions will also be shared through continuing collaboration with non-profit organizations with footprints in Africa. The research effort will be incorporated into educational activities for graduate students, symposium and workshop attendants, and high-schoolers via their teachers.This project is funded by the Hydrologic Sciences program, as well as a collaboration between the Directorate for Geosciences and Office of Advanced Cyberinfrastructure to support AI/ML and open science activities in the geosciences.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.