Orateur
Description
Numerical and experimental approaches are becoming more complementary approaches to address highly turbulent regimes, thanks to improvements in computational resources. Nonetheless, experimental results, such as meshless Particle Tracking Velocimetry, are often spatially sparse, somewhat discontinuous, subject to measurement noise and incomplete, while CFD simulations still require intensive resources. To address these issues, we employ Physics-Informed Neural Networks (PINNs), which provide a meshless way to enrich and complement experimental data.
In the case of turbulent convection, our framework allows 3D temperature discovery, denoising, and generating continuous field representations. We propose a spatio-temporal sampling framework to tailor our DNS database to closely resemble experimental measurements. This poses additional challenges to PINNs, particularly in addressing spatial gaps, tracking loss and scarcity of experimental-type labels. Through parallel GPU computations, our analysis focuses on the impact of PDE collocation points density and the effectiveness of smart adaptive spatial PDE sampling.
