EAGER: Scalable Climate Modeling using Message-Passing Recurrent Neural Networks

真实世界的气候模型通常严重依赖大规模的物理驱动的气候模型，涉及数百万未知参数和稀疏的真实世界测量，以准确校准这些模型。该提案旨在开发消息传递递归神经网络（MPRNN），这是一种深度图神经框架，用于从稀疏时空传感器测量中进行准确，可扩展和高效的气候建模。与细粒度的基于物理的气候模型不同，mprnn利用在不同空间节点上建立的离散、分布式异构循环神经网络集合，模拟基于物理的底层模型，并使用消息传递算法进行通信，以生成实时时空气候图。本提案对包括计算机科学、复杂系统和气候科学在内的几个子学科做出了根本性贡献，包括：(i)设计可扩展的图形机器学习框架，用于模拟复杂气候系统；（ii）使用消息传递算法和图形神经网络模拟基于物理的偏微分方程的底层时空模型；（iii）设计一个通用库，用于有效地实现基于MPRNN的复杂气候系统模拟器。该提案旨在证明MPRNN在多种气候建模工作中的有效性，包括重力波建模、环境污染预测和理解密集城市环境中气候变化的局部影响。该建议建立在研究人员先前的工作基础上，该工作为污染预测和重力波建模提供了MPRNN的基线实现。这项工作对气候研究的各个利益相关者有更广泛的影响，包括气候建模研究人员和解决气候变化、空气污染和重力波等重要问题的政策专家。通过将基于物理的域结构整合到深度图模型中，该建议可以使气候专家能够使用可扩展的基于深度图的气候模型有效地模拟大型基于物理的模拟模型的行为。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

Real-world climate models have often been heavily reliant on large-scale physics-driven climate models involving millions of unknown parameters and sparse real-world measurements to accurately calibrate these models. This proposal aims to develop Message Passing Recurrent Neural Networks (MPRNN), a deep graph neural framework for accurate, scalable and efficient climate modeling from sparse spatio-temporal sensor measurements. Unlike fine-grained physics based climate models that model continuous behavior across space and time, MPRNNs leverage a discrete, distributed collection of heterogeneous recurrent neural networks established at different spatial nodes that simulate the underlying physics-based model and communicate using message passing algorithms to generate real-time spatio-temporal climate maps. This proposal makes fundamental contributions across several sub-disciplines including computer science, complex systems, and climate sciences including: (i) designing scalable graph machine learning frameworks for modeling complex climate systems, (ii) simulating underlying partial differential equations based physics-informed spatio-temporal models using message passing algorithms and graph neural networks, and (iii) designing a general purpose library for efficiently implementing MPRNN based simulators for complex climate systems. This proposal aims to demonstrate the efficacy of MPRNN on multiple climate modeling efforts including modeling gravity waves, environmental pollution forecasting and understanding the localized impact of climate variations in dense urban environments. This proposal builds upon prior work by the investigators that provides a baseline implementation of MPRNN for pollution forecasting and gravity wave modeling. This line of work has a broader impact on various stakeholders of climate research, including climate modeling researchers and policy experts in tackling important issues like climate change, air pollution and gravity waves. By incorporating physics-based domain structure into deep graph models, this proposal can enable climate experts to effectively emulate the behavior of large physics based simulation models using scalable deep graph based climate models.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.