Title:Standardized Non-Intrusive Reduced Order Modeling Using Different Regression Models With Application to Complex Flow Problems

Abstract:In recent years, numerical methods in industrial applications have evolved from a pure predictive tool towards a means for optimization and control. Since standard numerical analysis methods have become prohibitively costly in such multi-query settings, a variety of reduced order modeling (ROM) approaches have been advanced towards complex applications. In this context, the driving application for this work is twin-screw extruders (TSEs): manufacturing devices with an important economic role in plastics processing. Modeling the flow through a TSE requires non-linear material models and coupling with the heat equation alongside intricate mesh deformations, which is a comparatively complex scenario. We investigate how a non-intrusive, data-driven ROM can be constructed for this application. We focus on the well-established proper orthogonal decomposition (POD) with regression albeit we introduce two adaptations: standardizing both the data and the error measures as well as -- inspired by our space-time simulations -- treating time as a discrete coordinate rather than a continuous parameter. We show that these steps make the POD-regression framework more interpretable, computationally efficient, and problem-independent. We proceed to compare the performance of three different regression models: Radial basis function (RBF) regression, Gaussian process regression (GPR), and artificial neural networks (ANNs). We find that GPR offers several advantages over an ANN, constituting a viable and computationally inexpensive non-intrusive ROM. Additionally, the framework is open-sourced to serve as a starting point for other practitioners and facilitate the use of ROM in general engineering workflows.