Reduced Order Modelling and Deep Neural Network-based Computational Method for Accelerating Parametric Investigation of Time-Periodic CFD Simulations

The present work is on development of a computational method, for accelerating parametric simulations of time-periodic unsteady Computational Fluid Dynamics (CFD) simulations. The proposed method starts with unsteady CFD simulations for a discrete input-parametric space, followed by Proper Orthogonal Decomposition (POD)-based data reduction of the periodic flow field (obtained from various combinations of parameters-based simulations), and finally using a Deep Neural Network (DNN)—on the reduced data—to accurately predict periodic (a) flow field and (b) engineering parameters on entire range of input parametric space. For the present CFD simulations, a level-set function-based immersed interface method (LS-IIM)-based in-house code is used [1]. Further, in-house code is developed for the ROM and DNN. Effectiveness of the ROM-DNN-based predictive hybrid CFD method is presented on two CFD problems: (a) flow across an elliptic cylinder for various upstream and downstream aspect ratios (AR_U=0 –1.0 and AR_D=0 – 1.0) at Reynolds number Re=100, and (b) flow induced vibration (FIV) on an elastically-mounted circular cylinder for various reduced velocity U*=3–12 at a constant mass ratio m*=2.0, damping coefficient z=0.005, and Re=100. For the CFD simulations, the discrete input-parametric space corresponds to fifteen various combinations of AR_U=0, 0.2, 0.6, and 1.0 and AR_D=0, 0.2, 0.6, and 1.0 for the former problem and six values of U*=3, 5, 6, 8, 10, and 12 are considered for the latter problem on FIV. The resulting data for the periodic flow-field, from the fifteen/six parameters at twenty-five discrete time-instants (within a time-period), is reduced by POD and thereafter, the DNN is used on the reduced flow-field data to predict the periodic flow field within the entire range of input parametric space. The present work is significant as the proposed ROM-DNN-based hybrid CFD method predict fluid dynamics results in the entire range of input parametric space, as compared to results at discrete parametric values by the traditional CFD method