Model Reduction for Advection Dominated Hyperbolic Problems in an ALE Framework: Offline and Online Phases

Davide Torlo
DOI: https://doi.org/10.48550/arXiv.2003.13735
2020-03-31
Abstract:Model order reduction (MOR) techniques have always struggled in compressing information for advection dominated problems. Their linear nature does not allow to accelerate the slow decay of the Kolmogorov $N$--width of these problems. In the last years, new nonlinear algorithms obtained smaller reduced spaces. In these works only the offline phase of these algorithms was shown. In this work, we study MOR algorithms for unsteady parametric advection dominated hyperbolic problems, giving a complete offline and online description and showing the time saving in the online phase. We propose an arbitrary Lagrangian--Eulerian approach that modifies both the offline and online phases of the MOR process. This allows to calibrate the advected features on the same position and to strongly compress the reduced spaces. The basic MOR algorithms used are the classical Greedy, EIM and POD, while the calibration map is learned through polynomial regression and artificial neural networks. In the performed simulations we show how the new algorithm defeats the classical method on many equations with nonlinear fluxes and with different boundary conditions. Finally, we compare the results obtained with different calibration maps.
Numerical Analysis
What problem does this paper attempt to address?
The main problem that this paper attempts to solve is the challenges encountered by model order reduction techniques when dealing with convection - dominated problems. Specifically: 1. **Model order reduction for convection - dominated problems**: Convection - dominated problems are very common in many engineering applications, such as shallow - water equations, gas - dynamics simulations or explosions, etc. These problems are typical in hyperbolic conservation laws because their weak solutions develop shocks from smooth initial data, and typical simulations involve shock propagation. Moreover, in convection - diffusion - reaction problems, such problems also occur when the convection parameter dominates the equation. 2. **Limitations of classical model order reduction methods**: Classical model order reduction (MOR) techniques are usually applicable to diffusion - dominated problems, but perform poorly in convection - dominated problems. This is because the information compression ability of convection - dominated problems is weak, resulting in a slow decay of the Kolmogorov N - width. This means that if one attempts to compress information in a linear subspace, such as a moving shock, the number of basis functions required is proportional to the degrees of freedom of the discrete domain. Therefore, classical MOR techniques are not suitable for such problems. 3. **Development of nonlinear algorithms**: In recent years, much research has been devoted to developing new nonlinear MOR techniques to deal with convection - dominated problems. These methods can be divided into two broad categories: Eulerian methods and Lagrangian methods. Eulerian methods adjust the reduction structure by adapting to a moving frame, while Lagrangian methods transform the solution before applying MOR techniques. 4. **MOR algorithms for the complete online and offline phases**: Although some research has already demonstrated the effectiveness of nonlinear MOR techniques in the offline phase, few studies provide a complete online and offline description, especially for parameterized time - dependent problems. This paper aims to fill this gap by proposing an arbitrary Lagrangian - Eulerian (ALE) method, which modifies the offline and online phases of the MOR process, aligns the convection characteristics at the same location, and thus significantly compresses the reduced space. 5. **Specific objectives**: - Propose a MOR algorithm based on the ALE framework, which is suitable for unsteady - state parameterized convection - dominated hyperbolic problems. - Provide a complete description of the offline and online phases of the algorithm and demonstrate the computational time savings in the online phase. - Use polynomial regression and artificial neural networks to learn calibration mappings for efficient computation. - Verify the superiority of the new algorithm on equations with nonlinear fluxes and different boundary conditions through numerical simulations. - Compare the results of different calibration mappings and evaluate their performance. In summary, the goal of this paper is to improve the model order reduction efficiency for convection - dominated problems by introducing a new ALE framework and an improved MOR algorithm, thereby significantly reducing the computational time in practical applications.