Improving Weather Forecasting through non-Gaussian Data Assimilation with Machine Learning

这项研究项目是为了提高使用数学学科将理论与观测最佳结合起来以提高天气预报精度的技术。该团队将使用不同形式的机器学习机制来检测大气中湿度场行为的变化，以便新技术能够改变天气预报方案的部分内容，以更好地捕获不同位置和不同时间的这些场。为了实现研究目标，需要大量的观测和模型结果来训练计算机检测水分变化。这项研究将调查需要多少数据才能通过机器学习技术可靠地检测变化。作为这一研究项目的一部分，研究团队将为研究界开发一个网站，以查看过去24小时内大气湿度的变化。这项研究还将测试当机器学习技术检测到湿度从正常行为中变化时，天气预报方案的新组成部分工作得如何。该项目还将包括培训一名新科学家学习最新的研究方法。研究小组将调查机器学习技术检测湿度场的高斯行为变化的能力，以及在变分数据同化中在高斯和非高斯之间切换成本函数的能力。该方案对于确保模型观测误差被一致地建模是重要的。误差的变化通常被认为是朝向对数正态的；最近的工作表明，水汽场的行为有另一个概率密度函数--反向对数正态。这种分布具有合适的偏斜度，并使分析能够在背景太干燥的情况下增加湿度状态。使用适当类型的误差分布方案，不仅有助于云预报，而且有助于在资料同化方案完成后云在预报模式中的保留。该团队还将研究一种新的集合平滑器，以及最大似然集合过滤器的非高斯版本，作为混合数据同化方案的更一致的集合过滤器，特别是对于对数正态和逆对数正态行为。此外，不同位置和高度的湿度场的偏斜度将显示在团队的网站上，供公众和天气预报员查看湿度分布的变化。这一奖项反映了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。

The research project is to advance techniques for using a mathematical discipline to optimally combine theory with observations to improve the accuracy of weather forecasts. The team will use different forms of machine learning mechanism to detect changes in the behavior of moisture fields in the atmosphere such that the new techniques are able to change parts of the weather prediction scheme to better capture these fields in different locations and at different times. To achieve the research goal, a large amount of observations and model results is required to train computers to detect moisture changes. The research will investigate how much data are needed to reliably detect changes through the machine learning techniques. As part of this research project, the research team will develop a website for the research community to view atmospheric moisture changes in the past 24 hours. This research will also test how well a new component of the weather prediction scheme works when the machine learning techniques have detected moisture changes from its normal behavior. The project will also involve training a new scientist to learn the latest research method. The research team will investigate the ability of machine learning techniques to detect changes away from Gaussian behavior for the moisture fields and to be capable to switch the cost function in variational data assimilation between Gaussian and non-Gaussian. The scheme is important to ensure that the model-observation errors are being model consistently. The error changes are commonly assumed to be toward lognormal; recent work has indicated that the behavior of moisture fields has another probability density function—the reverse lognormal. This distribution has a right skewness and enables analysis to increase the moisture state if the background is too dry. Using the proper type of error distribution schemes will aid not only cloud prediction but also cloud retention in forecast models after the data assimilation scheme has finished. This team will also investigate a new ensemble smoother, as well as non-Gaussian versions of the Maximum Likelihood Ensemble Filter as a more consistent ensemble filter for hybrid data assimilation schemes, especially for the lognormal and reverse lognormal behavior. In addition, the skewness of the moisture field at different locations and heights will be displayed at the team’s website for the general public and forecasters to view how the moisture distribution is changing.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.