Investigation of Deep Learning for Solid and Fluid Simulations
Abstract
Many problems within science and engineering rely upon numerical simulation to predict and understand the behavior of complex physical systems. Despite recent increases in available computational resources which have allowed researchers to begin considering these types of problems, the algorithms and models often struggle to closely match real-world data. We survey machine learning and deep learning, the latter of which is a promising methodology fordeveloping data-driven algorithms that can address these issues. The rise of deep learning has created newfound interest in continuous and applied mathematics problems such as nonlinear optimization and function interpolation, and we note that our lab is especially well-suited to address these topics, especially in the context of simulation. Building on our lab~s history ofnumerical algorithm design and development, the PI proposes to investigate the fusion of simulation and deep learning in order to innovate new numerical methods that aggressively leverage both disciplines and produce state-of-the-art results. We include the first real proposal to actually use deep learning as a core algorithmic component of a computational fluid dynamics simulation, as opposed to more superficial uses of learning; as such, the work stemming fromthis proposal might open an entire new field of cross-pollinating algorithm design.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- May 23, 2019
- Source ID
- N000141912285
Entities
People
- Ronald Fedkiw
Organizations
- Office of Naval Research
- Stanford University
- United States Navy