Understanding our physical world is clearly critical and beneficial to human society, which has become a central focus and challenge in many areas of science and engineering for centuries. This project will develop machine learning-based techniques to model complex atmospheric systems (from weather to climate). Atmospheric system models can approximate atmospheric flow and predict sequence of extreme precipitation events including flooding. Flooding is one the most deadly and costly natural hazards in the world. Mounting losses from catastrophic floods are driving an intense effort to increase preparedness and improve response to disastrous flood events by providing early warnings. Findings in this project will help decision makers better determine the need for and outcomes of particular policy actions. For example, a 10-15 day lead time in flood prediction will allow significant changes in the way reservoir operation rules are executed to minimize the impact of flood events. Moreover, this project will provide undergraduate and graduate students with valuable research and training opportunities, encourage minority and woman participation in science and engineering, and have a broad and sustainable impact on Computer Science curricula and courseware development.
Many physical systems can be described by a set of governing partial differential equations. However, these underlying governing partial differential equations are often coupled and nonlinear, do not have tractable analytical solutions, and need numerical approximations that are highly sensitive to initial and boundary conditions. This project synthesizes current understanding of physical systems with novel neural architectures to develop deep neural network models that can improve interpretation, generalization and prediction of complex physical system models. To achieve this goal, this project focuses on three interrelated research activities: (1) developing a library of neural architectural components to build modular neural network models; (2) testing neural architectural component based deep learning approach for flood prediction; and (3) building physics inspired deep learning models for better interpretation and prediction. This project investigates a new approach of developing and using basic neural architectural components to build large physics-informed deep neural networks. The modularity-based approach on study of neural architectures is critically important to enhance understanding and interpretability of deep learning models and has broad applications in multiple scientific domains. From scientific perspective, it will provide a new benchmark on the efficacy of using neural architectural components to build physics-informed deep neural network models and quantify achievable predictability limits for a class of precipitation and flood events by combining strengths of partial differential equation based numerical weather prediction models and recent advances in deep learning.
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.
