Understanding and modeling atmospheric subgrid processes with deep learning

面对全球变暖，迫切需要准确的气候预测。然而，当前的气候模式仍然存在很大的不确定性，这主要是由小于模式网格尺度的云的近似表示引起的，也称为参数化。云过程及其与湍流和辐射的相互作用是高度混乱的，这使得传统的基于物理直觉和手动调整的参数化发展缓慢而繁琐。最近，已经有可能运行全球高分辨率模拟，明确地解决复杂的云过程。然而，这些模拟的计算成本很高，这将它们的预测范围限制在至多几个月。然而，这种短期数据集可以被用来开发更好的气候模型参数。在这个项目中，深度学习将被用于系统地利用短期高分辨率模拟来进行气候模拟。深度学习是人工智能的一个分支，它基于多层人工神经网络，可以学习复杂的非线性关系。2018年，包括申请人在内的第一批研究证明了使用高分辨率数据为气候模型建立深度学习次网格参数化的总体可行性。拟议工作的一个关键目标是使用尖端机器学习技术来提高这些早期研究的数值稳定性和物理一致性，以便进行现实的气候模拟。另一个关键目标是使用深度学习参数来了解子网格过程本身。深度学习已被证明非常适合于模拟具有空间和时间结构的过程，这些过程在大气中也很重要，但没有包括在大多数当前的参数化中。提出的深度学习方法允许探测丰富的高分辨率数据集的空间和时间结构以及过程交互的重要性。建立更好的气候模型需要更有效地使用数据。深度学习提供了这样做的一种方式。拟议的工作计划旨在开发基本的方法，以便用机器学习的方法实现真实的气候模拟，但也要使用这一新技术从有史以来最详细的模拟中提取洞察力。此外，该项目将有助于弥合尖端机器学习研究和气候建模之间的差距。

In the face of global warming accurate climate predictions are urgently needed. Current climate models, however, still suffer from large uncertainties, which are mainly caused by the approximate representation, also called parameterization, of clouds smaller than the model grid scale. Cloud processes and their interactions with turbulence and radiation are highly chaotic, making traditional parameterization development, based on physical intuition and manual tuning, slow and cumbersome. Recently, it has become possible to run global high-resolution simulations which explicitly resolve complex cloud processes. These simulations are computationally expensive, however, which limits their prediction horizon to a few months at most. Nevertheless, such short-term datasets could be exploited to develop better parameterizations for climate models. In this project, deep learning will be used to systematically leverage short-term high-resolution simulations for climate simulations. Deep learning, a branch of artificial intelligence, is based on multi-layered artificial neural networks, which can learn complex nonlinear relations. In 2018, first studies, including the applicant's, have demonstrated the general feasibility of building a deep learning subgrid parameterization for climate models using high-resolution data. One key objective of the proposed work is to use cutting-edge machine learning techniques to improve the numerical stability and physical consistency of these early studies in order to run realistic climate simulations. Another key goal is to use deep learning parameterizations to learn about the subgrid processes themselves. Deep learning has proven to be an excellent fit for modeling processes with spatial and temporal structure, which are also important in the atmosphere but are not included in most current parameterizations. The proposed deep learning approach allows probing the rich high-resolution dataset for the importance of spatial and temporal structures, as well as process interactions. Building better climate models requires a more efficient use of data. Deep learning provides one way of doing so. The proposed work plan aims to develop essential methodology in order to achieve real climate simulations with a machine learning approach but also to use this novel technology to extract insight from the most detailed simulations ever available. Further, this project will help to bridge the gap between cutting-edge machine learning research and climate modeling.