Neural Network-Based Discrepancy Modelling of Reduced Order Models With Surface-to-Surface Radiation

Abstract

Accurate and computationally efficient models are critical for real-time applications and system-level simulations. Finite element method (FEM)-based models offer highly accurate physical representations but their complexity renders them unsuitable for real-time computations on inexpensive hardware. Projection-based model order reduction (MOR) techniques can alleviate this issue by simplifying FEM models while retaining much of their accuracy. However, their effectiveness varies significantly for nonlinear problems, and their intrusive nature presents challenges, particularly when commercial software is employed. This paper introduces a hybrid modelling approach that combines a reduced order model (ROM), derived from a readily available linear representation of the system, with corrections provided by an artificial neural network (ANN) trained on data easily collected from the non-linear representation. The proposed method is applied to develop a lightweight thermal model of a power converter, capable of accurately reconstructing temperature distributions while accounting for non-linear surface-to-surface and surface-to-ambient radiation effects.