AGS-PRF: Understanding Tropical High Cloud Feedbacks via Machine Learning and Super Parameterization

对流降雨通常以深积云的形式出现，在雷达图像上显示为明亮的红色斑点，嵌入更大的较弱降雨区域，颜色为橙色或黄色。这些降雨量较弱的大片地区通常含有高度较高、相对平坦的云层，类似于上图所见的孤立雷暴的“砧板”。高层状云是对流单体的结果，就像只有在雷雨充分发展后才会形成砧板一样。但大范围高空层云的存在会影响后续对流云的发展，也就是促进大范围的对流聚集和组织。高层云通过阻挡向外发射的红外线辐射，加热下面的柱体，促进上升运动，从而鼓励新的云形成。这种影响被认为在热带海洋上很重要，因为在热带海洋中，组织高纬度对流的锋面天气系统基本上是不存在的。但对这种红外云辐射反馈及其对热带天气和气候的影响的了解目前相当有限。该奖项下的工作涉及云辐射反馈在马登-朱利安振荡(MJO)发展中的作用，在MJO中，大片对流区在热带印度洋上空组织，并缓慢向东传播30至60天。云辐射反馈是MJO对流组织的一个原因，也是决定其传播速度的一个关键因素。该提案涉及的第二个问题是对流组织在确定地球气候敏感性方面可能发挥的作用，即温室气体浓度在给定的增减中导致的全球变暖或变冷的程度。这项研究是通过观测分析和数值模拟相结合的方式进行的。观测分析使用应用于卫星和气象气球数据的人工神经网络(ANN)来确定层云发展与环境温度、湿度和云层稳定性的关系。使用分层相关传播(LRP)来解释由ANN识别的关系，LRP是一种识别对于生成ANN结果最重要的输入的技术。像人工神经网络这样的机器学习工具对基础科学的价值经常受到质疑，因为它们的“黑匣子”性质：尽管机器学习方法可以具有不可思议的预测能力，但它们的结果并没有解释为什么一组特定的输入会产生给定的结果。因此，LRP是一种打开黑匣子并从通过ANN机制确定的经验关系中获得物理洞察和理解的手段。由于MJO的全球影响，MJO影响着天气和气候现象，包括热带气旋(特别是在墨西哥湾)、季风降雨量和时间以及厄尔尼诺事件的发生，因此这项工作具有社会意义。由于MJO事件传播缓慢，原则上是可以预测的，但在当前的天气和气候模式中，预测技能是有限的。考虑到温室气体浓度的迅速上升，对流组织对气候敏感性的可能影响也是社会关注的问题。该项目还通过为职业生涯早期的科学家提供支持来促进劳动力发展。该奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

Convective rainfall often takes the form of deep cumulus clouds, which appear as bright red spots on a radar image, embedded in a larger region of weaker rainfall colored orange or yellow. These broad areas of weaker rainfall typically contain high, relatively flat clouds similar to the "anvils" seen above isolated thunderstorms. The high stratiform clouds are a consequence of the convective cells, just as the anvil forms only after a thunderstorm is fully developed. But the presence of a broad region of upper-level stratus can affect the subsequent development of convective clouds, or in other words it can promote large-scale convective aggregation and organization. Upper-level clouds encourage new cloud formation by blocking outgoing infrared radiation, heating the column below and promoting rising motions. The effect is thought to be important over tropical oceans where the frontal weather systems that organize convection in higher latitudes are largely absent. But understanding of this infrared cloud-radiation feedback and its effects on tropical weather and climate is currently quite limited.Work under this award addresses the role of cloud-radiation feedback in the development of the Madden-Julian Oscillation (MJO), in which a large region of convection organizes over the tropical Indian Ocean and propagates slowly eastward for a period of 30 to 60 days. Cloud-radiation feedback has been invoked as a cause of convective organization in the MJO and as a key factor in determining its propagation speed. A second issue addressed in the proposal is the possible role of convective organization in determining the Earth's climate sensitivity, meaning the amount of global warming or cooling that results from a given increase or decrease in greenhouse gas concentrations.The research is conducted through a combination of observational analysis and numerical simulations. The observational analysis uses artificial neural networks (ANNs) applied to satellite and weather balloon data to determine how stratus cloud development relates to ambient temperature, moisture, and stability in the cloud layer. The relationships identified by the ANN are interpreted using layer-wise relevance propagation (LRP), a technique that identifies the inputs which matter the most for generating an ANN result. The value of machine learning tools like ANN for basic science is often questioned because of their "black box" nature: while machine learning methods can have uncanny predictive power, their results do not come with any explanation for why a particular set of inputs produces a given result. LRP is thus a means to open the black box and gain physical insight and understanding from the empirical relationships identified through the ANN machinery.The work has societal relevance due to the worldwide effects of the MJO, which influences weather and climate phenomena including tropical cyclones (particularly in the Gulf of Mexico), the amount and timing of monsoon rainfall, and the onset of El Nino events. MJO events are predictable in principle given their slow propagation, but prediction skill is limited in current weather and climate models. The possible influence of convective organization on climate sensitivity is also of societal interest given the rapid rise of greenhouse gas concentrations. The project also contributes to workforce development by providing support to an early-career scientist.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.