Sandia National Labs FY20 LDRD Annual Report


Integrating state-of-the-art machine learning technologies with multiscale modeling frameworks. Projection-based reduced-order models (ROMs) provide a promising approach for enabling accurate, low-cost simulations of dynamical systems for applications such as engineering design and uncertainty quantification. Unfortunately, state-of-the-art ROM approaches can become inaccurate when applied to systems characterized by nonlinear, nonsymmetric, and multiscale dynamics. In this research, Sandia developed a novel ROM methodology — termed windowed least-squares (WLS) — to address these shortcomings. WLS leverages ideas from space-time model reduction and residual minimization to develop a ROM approach capable of providing accurate and efficient reduced-order models of complex dynamical systems. The methodology was additionally integrated into an open-source reduced-order modeling code developed at Sandia, termed

Pressio, and the approach demonstrated on benchmark supersonic and geophysical fluid dynamics problems. The research article, “Windowed least-squares model reduction for dynamical systems,” published in the Journal of Computational Physics , provides more details. (PI: Eric J. Parish) Numerical schlieren of supersonic cavity flow reduced-order model simulations. The new WLS method is closer to the truth solution than other state-of-the-art methods such as least- squares Petrov-Galerkin.

Adam Backer, an optical scientist at Sandia, is pushing technology’s limits to observe a fundamental feature of stretched DNA. (Photo by Randy Montoya)



Made with FlippingBook - professional solution for displaying marketing and sales documents online