Speaker
Description
The construction of reliable surrogate models for turbulent flows remains a major challenge in scientific machine learning. While projection-based Reduced Order Models (ROMs) provide mathematically grounded low-dimensional representations of fluid systems, standard Galerkin approaches often fail to produce physically consistent reduced turbulence closures.
In this work, we propose a hybrid surrogate modeling strategy that combines structure-preserving projection with data-driven learning. A discretize-then-project POD–Galerkin framework is employed to approximate velocity and pressure fields of the incompressible Navier–Stokes equations discretized via a finite-volume consistent flux method on collocated grids. To overcome the limitations of intrusive projection for turbulence modeling, the turbulent viscosity is instead reconstructed through a non-intrusive neural closure.
The mapping between the reduced velocity–pressure dynamics and the turbulent viscosity coefficients is learned using recurrent and attention-based neural architectures. A comparative study between Multilayer Perceptrons, Transformers, and Long Short-Term Memory networks shows that recurrent modeling significantly improves temporal stability and predictive accuracy in convection-dominated regimes.
Numerical experiments on a three-dimensional lid-driven cavity demonstrate that the proposed hybrid surrogate retains the physical consistency of projection-based ROMs while enhancing robustness in turbulent settings. The results illustrate how combining physics-based reduction with data-driven closure mechanisms can yield stable and accurate reduced-order surrogates for complex fluid systems.