Reconstruction of numerical inlet boundary conditions using machine learning: Application to the swirling flow inside a conical diffuser

• Published in: Physics of fluids (1994) 2021-08, Vol.33 (8)
• Main Authors: Véras, Pedro, Balarac, Guillaume, Métais, Olivier, Georges, Didier, Bombenger, Antoine, Ségoufin, Claire
• Format: Article
• Language: English
• Subjects: Boundary conditions; Computational fluid dynamics; Conical flow; Core flow; Diffusers; Engineering Sciences; Flow separation; Fluid dynamics; Fluids mechanics; Kinetic energy; Large eddy simulation; Machine learning; Mechanics; Numerical prediction; Physics; Simulation; Swirling; Turbulence; Turbulent flow; Velocity distribution
• ISSN: 1070-6631; 1089-7666
• DOI: 10.1063/5.0058642

A new approach to determine proper mean and fluctuating inlet boundary conditions is proposed. It is based on data driven techniques, i.e., machine learning approach, and its goal is to use any known information about the downstream flow to reconstruct the unknown or incomplete inlet boundary conditions for a numerical simulation. The European Research Community On Flow, Turbulence And Combustion (ERCOFTAC) test case of the swirling flow inside a conical diffuser is investigated. Despite its relatively simple geometry, it constitutes a very challenging test case for numerical simulations due to incomplete experimental data and to the delicate balance between core flow recirculation and boundary layer separation. Simulations are performed using both Reynolds averaged Navier–Stokes (RANS) and large-eddy simulations (LES) turbulence methods. The mean velocity and turbulence kinetic energy profiles obtained with the machine learning approach in RANS are found to be in very good agreement with the experimental measurements and the numerical predictions are greatly improved as compared to the previous results using basic inlet boundary conditions. They are indeed comparable to the best previous RANS using empirical ad hoc inlet conditions to accurately simulate the downstream flow. In LES, in addition to the mean velocity profiles, the machine learning approach also allows us to properly reconstruct the fluctuating part of the turbulent field. In particular, the methodology allows us to circumvent the lack of turbulent correlations associated with classical inlet synthetic turbulence.