Comparison of Applying FOUR Reduced Order Models to a Global Sensitivity Analysis

Abstract:

This study focuses on the comparison of applying four reduced order models (ROMs) to global sensitivity analysis (GSA). ROM is one way to improve computational efficiency in many-query applications such as optimization, uncertainty quantification, sensitivity analysis, inverse modeling where the computational demand can become large. The four ROM methods are: arbitrary Polynomial Chaos (aPC), Gaussian process regression (GPR), cut high dimensional model representation (HDMR), and random sample HDMR. The discussion is mainly based on a global sensitivity analysis performed for a hypothetical large-scale CO₂ storage project. Pros and cons of each method will be discussed and suggestions on how each method should be applied individually or combined will be made.