Development of a Physics-Data Driven Surface Flux Parameterization for Flow in Complex Terrain

用于天气预报和大气研究的数值模型是非常复杂的，但它们仍然必须对大气过程做出一些广义的假设，以提高计算效率。地球表面和大气之间的质量、能量和动量的交换是通过一个总体理论来评估的，该理论适用于平坦和均匀的地形，但不太适用于复杂的地形和可变的表面。在这个项目中，研究团队将开发和研究一种新的机器学习模型，以应对准确表征地表-大气交换的挑战。超过70%的地球陆地表面是复杂地形，提高天气和气候模式在这些地区的预测能力将有利于天气预报、野火控制、航空和军事应用。此外，该项目还有几个活动，旨在为学生提供探索物理和机器学习之间交集的能力。在过去的几十年里，Monin-Obukhov相似理论（MOST）一直是天气预报和气候预测模式中评估地球表面与大气之间质量、能量和动量交换的主要方法。然而，当应用于复杂的地形环境时，MOST存在着众所周知的缺陷。该项目将有助于开发一种基于物理的神经网络（PINN）模型，该模型有望提供更准确的面积聚集表面通量估计，并能够更直接和更合理地同化稀疏观测值以进行参数估计。该项目的第一步是通过一套过程分辨大涡模拟（LES）来生成复杂地形中微尺度流动的数值数据库。然后，该数据库将用于训练基于物理的表面通量神经网络（PINN-FLUX）。该项目的最后一项任务将是评估PINN-FLUX利用稀疏的原位观测来评估复杂地形中地表通量的能力。评估任务将通过与建模和观测的比较来进行。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。

Numerical models that are used for weather forecasting and atmospheric research are extraordinarily complex, yet they still must make some generalized assumptions about atmospheric processes to be computationally efficient. The exchange of mass, energy, and momentum between the earth’s surface and atmosphere is evaluated by an overarching theory that works well for flat and homogeneous terrain, but less so for complex terrain and variable surfaces. In this project, the research team will develop and investigate a new machine-learning model to tackle the challenge of enabling accurate characterization of surface-atmosphere exchange. More than 70% of Earth’s land surface is in complex terrain, and improving on the ability of weather and climate models projections in these areas will be beneficial for weather forecasting, wildfire control, aviation, and military applications. Additionally, the project has several activities that are intended to provide students with the ability to explore the intersection between physics and machine learning.The Monin-Obukhov Similarity Theory (MOST) has served as the primary method for evaluating the exchange of mass, energy, and momentum between the Earth's surface and the atmosphere in weather forecasting and climate projection models over the past decades. However, MOST has well-known deficiencies when applied to complex terrain environments. This project will enable the development of a physics-informed neural network (PINN) model that is expected to provide more accurate estimates of area-aggregate surface fluxes and enable a more straightforward and physically-justified assimilation of sparse observations for parameter estimation. The initial step in the project is the generation of a numerical database of microscale flow in complex terrain via a suite of process-resolving Large Eddy Simulations (LES). This database will then be used to train the physics-informed neural network for surface fluxes (PINN-FLUX). The final task in the project would be an assessment of PINN-FLUX’s ability to evaluate surface fluxes in complex terrain making use of sparse in-situ observations. The assessment task will be conducted using comparisons to modeling and observations.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.