Collaborative Research: Physics-Based Machine Learning for Sub-Seasonal Climate Forecasting

点击翻译按钮获取中文摘要

While the past few decades have seen major advances in weather forecasting on time scales of days to about a week, making high quality forecasts of key climate variables such as temperature and precipitation on sub-seasonal time scales, the time range between 2 weeks and 2 months, continues to challenge operational forecasters. Skillful climate forecasts on sub-seasonal time scales would have immense societal value in areas such as agricultural productivity, hydrology and water resource management, transportation and aviation systems, and emergency planning for extreme events such as Atlantic hurricanes and midwestern tornadoes. In spite of the scientific, societal, and financial importance of sub-seasonal climate forecasting, progress on the problem has been limited. The project has initiated a systematic investigation of physics-based machine learning with specific focus on advancing sub-seasonal climate forecasting. In particular, this project is developing novel machine learning (ML) approaches for sub-seasonal forecasting by leveraging both limited observational data as well as vast amounts of dynamical climate model output data. Further, the project is focusing on improving the dynamical climate models themselves based on ML with specific emphasis on learning model parameterizations suitable for accurate sub-seasonal forecasting. The principles, models, and methodology for physics-based machine learning being developed in the project will benefit other scientific domains which rely on dynamical models. The project is establishing a public repository of a benchmark dataset for sub-seasonal forecasting to engage the wider data science community and accelerate progress in this critical area. The project is training a new generation of interdisciplinary scientists who can cross the traditional boundaries between computer science, statistics, and climate science.The project works with two key sources of data for sub-seasonal forecasting: limited amounts of observational data and vast amounts of output data from dynamical model simulations, which capture physical laws and dynamics based on large coupled systems of partial differential equations (PDEs). The project is investigating the following central question: what is the best way to learn simultaneously from limited observational data and imperfect dynamical models for improving sub-seasonal forecasts? The project is building a framework for physics-based machine that has two inter-linked components: (1) deduction, in which ML models are trained on dynamical model outputs as well as limited observations, and (2) induction, in which ML models are used to improve dynamical models. Across the two components, the project is making fundamental advances in learning representations, functional gradient descent, transfer learning, derivative-free optimization and multi-armed bandits, Monte Carlo tree search, and block coordinate descent. On the climate side, the project is building an idealized dynamical climate model and doing an in depth investigation on learning suitable parameterizations for the dynamical model with ML methods to improve forecast accuracy in the sub-seasonal time scales. This project is part of the National Science Foundation's Harnessing the Data Revolution (HDR) Big Idea activity.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.