Data-Driven Model Reduction, Scientific Frontiers, and Applications ()
- Texas A&M University
- College Station, TX
- Joe C. Richardson Petroleum Engineering Building (RICH) 910
Freddie Witherden, Department of Ocean Engineering
Online Compression of High-Order CFD Solutions Using Machine Learning
Abstract
Unsteady computational fluid dynamics (CFD) simulations have the potential to generate petabytes of solution data; well beyond what can be handle by current generation I/O and storage sub-systems. As such, it is typically not practical to write out a sequence of finely spaced solution snapshots. Instead, the user must instrument the simulation in advance. However, this requires the user to have some a priori knowledge about the dynamics and evolution of the system—something which negates many of the advantages inherent to the simulation. In this talk, I will show how, within the context of high-order methods, a combination of convolutional autoencoders, principal component analysis, and interpolation can be employed to perform online lossy compression of unsteady CFD simulations. Such compression has the potential to enable routine snapshotting of even the largest simulations, thus facilitating far richer a posteriori analysis. Numerical experiments showcasing a five-fold reduction in archival storage requirements will be presented and the resulting flow fields will be compared and contrasted against the reference data.