Using artificial intelligence and inexact computing to improve modeling of multi-scale, multi-physics, chaotic dynamical systems with applications to weather/climate predictions
Abstract
There is an increasing need for improved modeling of complex, large-scale, multi-physics natural and engineering systems. Such systems are often described by multi-scale, high-dimensional, nonlinear partial differential equations (PDEs), solving which requires enormous computing resources. Accurate prediction of climate variability and weather extremes is a paramount example of this challenge. While the demand for better models/predictions keeps rising, the conventional frameworks for modeling such systems have reached their limits, and improvements in traditional computing hardware have substantially slowed down. Recently, artificial intelligence (AI) and inexact computing have received much attention, separately, as potential avenues for groundbreaking advances that enable increasing accuracy without requiring more computing resources. In the AI-based approach, data-driven surrogate models for too-expensive-to-resolve or poorly-understood processes are developed. The idea of inexact computing is that in complex systems, sometimes the precision of floating-point arithmetic can be reduced (which can substantially accelerate computations) without loss of accuracy because rounding error is dwarfed by other types of errors, e.g. from measurements or model biases. Explorations have shown that both approaches can reduce the computational cost and/or improve the accuracy of toy test models. However, various challenges remain before these approaches can be used for real-world problems. The objective of this proposal is to address some of these challenges by replacing ad hoc approaches with rigorous and physics/PDE-guided approaches, and also to develop a novel synergically integrated AI+Inexact computing framework.In the proposed work, novel ways of using AI-based data-driven techniques for modeling of multi-scale, multi-physics, chaotic systems are developed by closely connecting the AI techniques with the underlying physics and PDEs. Then, by deriving a theoretical model for how different types of error grow in geophysical turbulence, a rigorous framework for inexact computing of geophysical turbulence is built. Finally, AI and inexact computing are integrated synergically to develop a novel computational framework. A hierarchy of chaotic models and prototypes of atmospheric/oceanic turbulence are used to test the scalability of the framework to higher dimensional and more complex systems. The projects expected outcomes are 1) Data-driven AI techniques that are tailored for climate/weather/turbulence modeling, and novel frameworks for utilizing them for multi-scale, multi-physics chaotic/turbulent systems, 2) Physics-guided insight into precision selection for applying inexact computing to geophysical turbulence, 3) A paradigmshifting computational framework that synergically integrates conventional modeling approaches with AI and inexact computing to provide unprecedented accuracy.These outcomes closely align with the ONRs interests in developing innovative mathematical and computational tools to model multi-scale, multi-physics dynamical systems such as atmosphere/ocean/ice, and to better understand and predict global climate oscillations and teleconnections, high-latitudes, and stratosphere. This works expected outcomes will enhance the DOD capabilities in environmental modeling and prediction for operational planning attactical/strategic scales and for increasing mission success. They also give the DOD a competitive advantage by harvesting the emerging power of AI, data, and novel computational frameworks for enhanced insight, rapid decision-making, and new mission capabilities.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jul 20, 2020
- Source ID
- N000142012722
Entities
People
- Pedram Hassanzadeh
Organizations
- Office of Naval Research
- Rice University
- United States Navy