Prediction of the dynamics of a backward-facing step flow using focused time-delay neural networks

Antonios Giannopoulos (1) , Jean-Luc Aider (1)

1 PMMH - Physique et mécanique des milieux hétérogenes (UMR 7636)

The full field dynamics of a separated, noise-amplifier flow, the Backward-Facing Step at Re h = 1385, have been identified by probe-like, upstream measurements using an artificial Neural Network. Local visual sensors, coming from time-resolved Particle Image Velocimetry, were used as inputs and the dynamic Proper Orthogonal Decomposition coefficients were defined as goals-outputs for this non-linear mapping. The coefficients time-series were predicted and the instantaneous velocity fields were reconstructed with satisfying accuracy. The choices of inputs-sensors, training data-set size, hidden layer neurons and training hyperparameters are discussed for this experimental fluid system.

Mots clés: Backward-facing step flow; Neural Networks; Particle Image Velocimetry; Machine learning; System Identification

Domaines: Mécanique des fluides [physics.class-ph]

Antonios Giannopoulos, Jean-Luc Aider. Prediction of the dynamics of a backward-facing step flow using focused time-delay neural networks. International Journal of Heat and Fluid Flow, 2020, 82, pp.108533. ⟨10.1016/j.ijheatfluidflow.2019.108533⟩. ⟨hal-03007152⟩