Orateur
Description
We study the dynamics of key quantities relevant to RANS modeling of the planar Rayleigh–Taylor instability, averaged along homogeneous directions (1D profiles). Our work relies on a database of 484 DNS of resolution 1048 x 2048 × 2048 points, spanning a wide range of initial conditions.
This dataset serves to build a generalized and interpretable Reduced Order Model (ROM) using Proper Orthogonal Decomposition (POD). The reduced representation is then used to train a physics-informed neural network (PINN) with the residual-based attention framework and SSBroyden optimizer, predicting the temporal evolution of the POD modes from the initial conditions. By combining POD spatial modes with the surrogate model for their temporal dynamics, we reconstruct the full spatio-temporal evolution from the initial conditions.
We seek analytical representations of the ROM via sparse identification of nonlinear dynamics and inference of initial conditions from observations to predict the system’s full evolution.