CAREER: New Frontiers In Large-Scale Spatiotemporal Data Analysis

这项为期五年的职业发展计划旨在建立一个协同的研究和教育计划，推动对大规模时空数据(即跨空间和跨时间收集的数据)的分析，朝着高效、可靠和可靠的实时决策方向发展。从气候科学到公共卫生等各个科学领域，海量时空数据正在迅速涌现。传统的时空分析工具要么依赖于强大的建模假设，要么速度太慢，无法实时运行。虽然深度学习提供了很大的灵活性和可扩展性，但它理解大规模时空数据并最终为科学领域做出贡献的能力是有限的。一个主要原因是时空数据的独特性质：它是高度动态的，受物理规律支配，具有错综复杂的相互作用。这些特点对现有的机器学习方法提出了根本的挑战。受物理科学用例的启发，本研究计划寻求开发动态学习技术，以解决三个核心挑战：(1)在符合物理规律的情况下预测时空动力学；(2)推断时空相互作用以获取复杂的依赖关系；(3)量化时空预测的不确定性以进行决策。最终目标是设计能够比数值求解器更快、更准确地模拟洋流、交通流量和疫情传播的DL工具，从而实现实时情景规划、控制和策略优化。该教育计划将开发本科生、研究生水平的新课程，以及大规模开放在线课程(MOOC)。外展活动将强调妇女和少数族裔早期参与机器学习研究。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。

This five-year career development plan aims to build a synergistic research and education program that advances analysis of large-scale spatiotemporal data (i.e., data collected across space and time) towards efficient, robust, and trustworthy real-time decision-making. Massive spatiotemporal data are emerging rapidly in various scientific fields from climate science to public health. Traditional spatiotemporal analysis tools either rely on strong modeling assumptions or are too slow to operate in real-time. While deep learning (DL) offers great flexibility and scalability, its ability to make sense of large-scale spatiotemporal data and ultimately contribute to scientific fields, is however, limited. A primary reason is the distinctive nature of spatiotemporal data: it is highly dynamic, governed by physical laws and has intricate interactions. These characteristics pose fundamental challenges to existing machine learning approaches.Inspired by the use cases in physical sciences, this research plan seeks to develop DL techniques that address three central challenges: (1) forecasting spatiotemporal dynamics while conforming to physical laws; (2) inferring spatiotemporal interactions to capture complex dependencies; and, (3) quantifying the uncertainty of spatiotemporal forecasts for decision making. The ultimate goal is to design DL tools that can emulate ocean currents, traffic flows and epidemic spread faster and more accurately than numerical solvers, thus allowing real-time scenario planning, control and strategy optimization. The education plan will develop new curricula at undergraduate, graduate level and massive open online courses (MOOCs). The outreach activities will emphasize the early engagement of women and minorities in machine learning 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.