Physics-Informed Neural Networks (PINNs)

Nowadays, the domain of fluid mechanics has seen significant advancements. Alongside traditional experimental datasets, Direct Numerical Simulations (DNS) of the Navier-Stokes equations have generated vast amounts of high-fidelity data​. This “wealth” of information has offered crucial insights for developing turbulence models, setting the stage for integrating data-based strategies into the classical turbulence modeling techniques. Classical machine learning (ML) and deep learning (DL) frameworks have emerged, opening new pathways towards constructing fast and reliable models that can replace and/or speed up traditional simulation methods and experimental settings. However, data-driven models may often face challenges such as artifacts, limited extrapolation ability, applicability, and explainability. Scientific machine learning, following the introduction of Physics-Informed Neural Networks (PINNs), represents a new approach that incorporates governing laws from physics, in order to numerically solve classical engineering problems. Partial differential equations (PDEs) are introduced into the loss functions of deep neural networks, usually combined with sparse data from experiments or simulations, and offer a mechanism to enforce physical laws consistency inside the traditional ML/DL approach. MOVEFREE research focuses on the application of PINNs in turbulent channel flows, where fluid dynamics laws are combined with sparse DNS and/or empirical equations data, in order to establish a computational framework to bypass resource- demanding DNS simulations with similar accuracy.