CFD stability improvement using dynamic mode decomposition of solution update vectors

Abstract

We present a novel method for improving the numerical stability of finite volume simulations by optimizing the mesh through dynamic mode decomposition of solution update vectors. Our approach leverages dynamic mode decomposition to approximate the non-linear solution evolution as a linear mapping, enabling us to represent the dynamic system with fewer degrees of freedom. By conducting eigenanalysis on the reduced system, we gain insights into the growth rate and oscillation frequency of dominant solution modes. We then identify the control volumes and vertices that have a significant influence on each dynamic solution mode. We compute the gradients of the Jacobian diagonal elements with respect to the movement of the selected vertices. Based on these gradients, we adjust the positions of the vertices to improve the diagonal dominance of the corresponding Jacobian rows. We utilize this approach in conjunction with our in-house flow solver as well as with Ansys Fluent to test its effectiveness when applied to different types of CFD software architecture. We show that the presented methodology is fully non-invasive to the host flow solver and does not require any modifications or access to the source code. Our approach offers a substantial computational cost reduction compared to existing methods for numerical stability improvement. Various results demonstrate the strength and efficacy of this state-of-the-art approach for improving numerical stability through mesh optimization.