Sandia National Labs FY20 LDRD Annual Report

FY20 ANNUAL REPORT

Project Highlights – Technical Vitality LDRD is essential to maintaining the Labs’ scientific vitality, and Sandia, as the nation’s most diverse national security laboratory, is uniquely equipped to tackle groundbreaking, interdisciplinary research. Researchers collaborate across a broad spectrum of disciplines and achieve research breakthroughs, which advance our scientific understanding, enable national security, and result in technology transfer to industry that is commercialized under licensing agreements and brought to market for the U.S. public good. Incorporating physical constraints into Gaussian process surrogate models. Surrogate models are widely used as emulators for expensive physics simulations, but previously no open source or commercial class of machine learning (ML) models, called Gaussian processes (GP), treat constraints. The incorporation of physical constraints and domain knowledge into ML algorithms is a significant challenge when it comes to the application of ML algorithms to scientific and engineering FY20 LDRD Annual Report Proj ct Highlights – Mission Agility, Technical Vitality, Workforce Dev. Final Projec Highlights – Technical Vitality LDRD is essential to maintaining the Labs’ scientific vitality, and Sandia, as the nation’s most diverse national security laboratory, is uniquely equipped to tackle groundbreaking, interdisciplinary research. Researchers collaborate across a broad spectrum of disciplines and achieve research breakthroughs, which advance our scientific understanding, enable national security, and result in technology transfer to industry that is commercialized under licensing agreements and brought to market for the U.S. public good. Incorporating physical constraints into Gaussian process surrogate models. Surrogate models are widely used as emulators for expensive physics simulations, but previously no open source or commercial class of machine learning (ML) models, called Gaussian processes (GP), treat constraints. The incorporation of physical constraints and domain knowl dge into ML algorithms is a significant challen e when it comes to the application of ML algorithms to scientific and engineering problems. Sandia researchers are investigating Gaussian processes and examined bound constraints, monotonicity constraints, linear operator constraints that represent physical laws as partial differential equations, and intrinsic boundary condition constraints. The resulting approaches are expected to serve as a foundation for physics- constrained Gaussian process applications. (PI: Laura Painton Swiler)

problems. Sandia researchers are investigating Gaussian processes and examined bound constraints, monotonicity constraints, linear operator constraints that represent physical laws as partial differential equations, and intrinsic boundary condition constraints. The resulting approaches are expected to serve as a foundation for physics-constrained Gaussian process applications. (PI: Laura Painton Swiler)

Commented [TA54]: AST - Swiler3.png

Effect of enforcing a boundary condition when developing Gaussian process surrogates to model the solution of a one-dimensional problem: Effect of enforcing a boundary condition when developing Gaussian process surrogates to model the solutio of a one-dimensional problem: − ! ! !# " ! = ( ) ∈ ( 0,1 ) , ( 0 ) = 0, ( 1 ) = 0 Five scatt red observations of f are provided at the points l with X. The l ft panel shows a Five scattered observations of f are provided at the points labeled with X. The left panel shows a formulation that incorporates the differential equation information into the GP, but the inference fails with only information about f and not u. The right panel shows the GP with incorporation of the boundary value problem and the improved prediction. form l i that incorporates the differential equatio information into t e GP, but the infere ce fails with only information about f and not u. The right panel shows the GP with incorporation of the boundary value problem and the improved prediction.

Commented [TA55]: SUPPLEMENTAL PHOTO: https://www.gettyimages.com/detail/photo/artificial- intelligence-concept-royalty-free-image/1186776215.

42

LABORATORY DIRECTED RESEARCH & DEVELOPMENT