AGS-PRF: Physics-Constrained Machine Learning-Based Models for Climate Simulations with Data Assimilation and Uncertainty Quantification

尽管最近的计算进步，气候模式仍然不能明确地解决关键的物理过程，如湍流、对流和云。这些未解决的过程必须通过参数化方案加以解释。这些方案无法模拟现实，阻碍了模式模拟捕捉各种观测现象的能力，导致模式偏差和不确定性。最近，基于机器学习的技术极大地改进了这些方案。然而，标准的基于机器学习的方法完全建立在理想化的计算模型上，而没有考虑到丰富的观测数据。因此，参数化方案固有的不准确性继续破坏气候模式的性能。该项目旨在改进基于机器学习的参数化方案，从而提高现有气候模型的性能。提案人将把观测和物理定律纳入机器学习技术，使参数化方案更加准确，并提供不确定性估计，以捕捉气候系统的混沌本质。这项工作将培养一名年轻的博士后学者。具体来说，该提案将利用卫星观测数据和高分辨率模拟来改进机器学习方案。物理约束将通过合并物理损失项或考虑特定的机器学习工具（如具有固定输出层的神经网络）来施加已知的物理约束。机器学习方案中的不确定性量化将使用集成学习或贝叶斯方法（如哈密顿蒙特卡罗采样方案）来实现。在这个项目中获得的知识可能会对整个气候科学产生影响，包括全球气候模型的发展和机器学习在气候科学大数据中的应用，以及复杂的多尺度和多物理世界系统的新型计算概率方法的发展。该奖项反映了美国国家科学基金会的法定使命，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。

Despite recent computational advancements, climate models still cannot explicitly resolve key physical processes like turbulence, convection, and clouds. These unresolved processes must be accounted for by parameterization schemes. The inability of these schemes to mimic reality has hindered the ability of model simulations to capture various observed phenomena, leading model biases and uncertainties. Recently, machine learning-based techniques have greatly improved these schemes. Nevertheless, the standard machine learning-based methods are built solely on idealized computational models without considering the wealth of observational data available. Thus, inherent inaccuracies of the parameterization schemes continue to undermine the performance of climate models. This project aims to improve machine-learning-based parameterization schemes and, thereby, enhance the performance of existing climate models. The proposer will incorporate observations and physical laws into machine learning techniques to make parameterization schemes more accurate and provide uncertainty estimates to capture the chaotic nature of the climate system. The proposed work will train a young postdoctoral scholar. Specifically, the proposal will utilize data from satellite observations and high-resolution simulations to improve machine learning schemes. Physics-constraints will be imposed either through the incorporation of a physical loss term or by considering specific machine learning tools such as Neural Networks with fixed output layer that strongly imposes the known physical constraint. Uncertainty quantification in machine learning schemes will be implemented using ensemble learning or Bayesian approaches such as Hamiltonian Monte Carlo sampling schemes. The knowledge gained in this project could have an impact across climate science, including the advancement of global climate models’ development and of machine learning application to Big Data in climate science, as well as the development of novel computational probabilistic methods for complex multi-scale and multi-physics real-world systems.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.