EAGER: North American Monsoon Prediction Using Causality Informed Machine Learning

这项研究旨在更好地了解西南地区的季风雷暴活动和降水。该项目创建了一个创新的机器学习工具，使用区域数值天气模型输出和卫星遥感数据（预测器）对已知的雷暴单元位置和雷达检测到的强度（目标）进行培训。该工具将被设计为提取预测因子和目标之间的重要基本关系，以帮助解释雷暴的发展和演变。经过密集的培训、验证和测试阶段，这些关系将被用来更好地预测西南部未来雷暴事件的时间、严重程度和位置。该工具将与国家气象局共享，以帮助预报员预测雷暴相关的危害，如大冰雹，山洪暴发或野火点燃。这一创新方法还将为改进业务气象和地球物理预报系统和指导科学领域研究提供一个框架，该项目开发了一个概率模型来预测北美季风（NAM）湿期的对流开始、降雨率和对流单体轨迹。对流的预测因子（例如，将从动态中尺度模型（高分辨率快速刷新，亚利桑那大学-天气研究预报模型）分析和预报中收集有关土壤湿度和地表温度的数据（相对湿度、对流可用势能、可降水量），并将其与新的卫星土壤湿度和地表温度观测数据相结合，以产生一个独特的预测工具。一种新的机器学习方法-因果关系知情学习-将被应用于确定最合适的预测进一步训练的神经网络，并深入了解对流启动和演变的过程。每小时的降水发生，自然和分类降雨率的预报将产生业务指导预报员和现场research.This奖项反映了NSF的法定使命，并已被认为是值得通过使用基金会的智力价值和更广泛的影响审查标准进行评估的支持。

This research seeks better understanding of monsoon thunderstorm activity and precipitation in the Southwest. The project creates an innovative machine learning tool trained using regional numerical weather model output and satellite remote sensing data (the predictors) with respect to known thunderstorm cell locations and intensities detected by radar (the targets). The tool will be designed to extract important fundamental relationships between the predictors and targets that help explain the development and evolution of thunderstorms. After an intense training, validation and testing phase, the relationships will then be leveraged to generate better forecasts of the timing, severity and location of future thunderstorm events in the Southwest. The tool will be shared with the National Weather Service to help forecasters predict thunderstorm-related hazards such as large hail, flash flooding or wildfire ignition. This innovative approach will also provide a framework for improving operational meteorological and geophysical prediction systems and for guiding scientific field studies.The project develops a probabilistic model to predict convective initiation, rain rates, and convective cell tracks during the wet phase of the North American Monsoon (NAM). Predictors of convection (e.g., relative humidity, convective available potential energy, precipitable water) will be collected from dynamic mesoscale model (High Resolution Rapid Refresh, University of Arizona-Weather Research Forecast model) analyses and forecasts and combined with new satellite-derived observations of soil moisture and surface temperature to produce a unique prediction tool. A novel machine learning approach – causality informed learning – will be applied to identify the most suitable predictors for further training in a neural network and to gain insight into the processes governing convective initiation and evolution. Hourly forecasts of precipitation occurrence, nature, and categorical rain rates will be produced operationally to guide forecasters and field research.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.