Blackbox Stencil Interpolation Method for model reduction

Abstract

Model reduction often requires modifications to the simulation code. In many circumstances, developing and maintaining these modifications can be cumbersome. Non-intrusive methods that do not require modification to the source code are often preferred. This thesis proposed a new formulation of machine learning, Black-box Stencil Interpolation Method, for this purpose. It is a non-intrusive, data-oriented method to infer the underlying physics that governs a simulation, which can be combined with conventional intrusive model reduction techniques. This method is tested on several problems to investigate its accuracy, robustness, and applicabilities.