Non-intrusive model reduction of advection-dominated hyperbolic problems using neural network shift augmented manifold transformations

Advection-dominated problems are commonly noticed in nature, engineering systems, and a wide range of industrial processes. For these problems, linear approximation methods (proper orthogonal decomposition and reduced basis method) are not suitable, as the Kolmogorov n-width decay is slow, leading to inefficient and inaccurate reduced order models. There are few non-linear approaches to accelerate the Kolmogorov n-width decay. In this work, we use a neural-network shift augmented transformation technique, that employs automatic-shit detection and detects the optimal non-linear transformation of the full-order model solution manifold M. We exploit a deep-learning framework to derive parameter-dependent bijective mapping between the manifold M and the transformed manifold M̃. It consists of two neural networks, 1) ShiftNet, to employ automatic-shift detection by learning the shift-operator, which finds the optimal shifts for numerous snapshots of the full-order solution manifold, to accelerate the Kolmogorov n-width decay, and 2) InterpNet, which learns the reference configuration and can reconstruct the field values of the same, for each shifted grid distribution. We construct non-intrusive reduced order models on the resulting transformed linear subspaces and employ automatic-shift detection for predictions. We test our methodology on advection-dominated problems, such as 1D travelling waves, 2D isentropic convective vortex and 2D two-phase flow test cases. This work leads to the development of the complete NNsPOD-ROM algorithm for model reduction of advection-dominated problems, comprising both offline-online stages.