Parametric Nonlinear Model Reduction Using Machine Learning on Grassmann Manifold with an Application on a Flow Simulation

This work introduces a parametric model order reduction (PMOR) approach that enhances an existing widely used technique based on proper orthogonal decomposition (POD) and discrete empirical interpolation method (DEIM) for parametrized nonlinear dynamical systems by employing machine learning procedures performed on a Grassmann manifold. In particular, distances between parameters are first computed based on a metric defined on the Grassmann manifold of solution spaces. Then, the distance information is utilized in the K-medoids clustering algorithm to partition parameters into classes with corresponding local solution spaces, which are further used to form a dictionary of local bases. The artificial neural network (ANN) is next used to build a classifier that can automatically identify the most suitable local basis from the dictionary for a given input parameter to construct a parametrized reduced-order model by the POD–DEIM approach. This work numerically demonstrates the significance of using distance on the Grassmann manifold of the solution spaces, instead of directly using the Euclidean distance on the parameter space. To validate the proposed method, numerical studies are performed on a parametrized 1D Burger’s equation and a viscous fingering in a horizontal flow through a 2D porous media domain. The proposed method is shown to have advantage in terms of accuracy when compared to the traditional global basis approach, as well as the local reduced-order basis approach based on the Euclidean metric.