CoSy-seminarium med Süleyman Cengizci

Datum
till
Plats
Ångströmlaboratoriet, Å2002
Typ
Seminarium
Föreläsare
Süleyman Cengizci
Arrangör
Matematiska institutionen
Kontaktperson
Jörgen Östensson

Süleyman Cengizci håller i ett CoSy-seminarium

Alla är välkomna och vi bjuder de 40 första som anmäler sig på en gratis lunchsmörgås.

Titel: Learning to Move the Mesh: Stabilization-Informed Neural Adaptation for Finite Element Simulations

Abstract: Accurate numerical simulation of convection-dominated transport remains challenging, particularly when sharp solution gradients and localized flow structures cannot be adequately resolved on computationally affordable meshes. Stabilized finite element methods improve numerical robustness, but their accuracy remains strongly dependent on the spatial distribution of computational resolution. In this seminar, I will present a neural mesh-adaptation framework that combines stabilized finite element methods with a pretrained Universal Mesh Movement Network (UM2N). The central idea is to exploit information already available within the numerical formulation, including solution curvature and local stabilization activity, to guide the redistribution of mesh vertices. By combining Hessian-based indicators with information derived from YZβ shock-capturing and LSIC stabilization, the proposed approach enables physics-guided mesh movement without modifying mesh connectivity or increasing the number of degrees of freedom. The neural network is employed in a zero-shot setting, without problem-specific retraining. Numerical examples involving convection-diffusion, viscous Burgers equations, Kovasznay flow, and unsteady incompressible flow past a circular cylinder demonstrate the capabilities and limitations of the approach. This is joint work with B. Emek Abali (Uppsala University).