EFFICIENT NON-INTRUSIVE MODEL ORDER REDUCTION FOR PARAMETRIC AEROELASTIC OBEJCTS USING CONVOLUTIONAL NEURAL NETWORKS BASED MACHINE LEARNING

A data-driven non-intrusive model order reduction (MOR) methodology for the parametrized aeroelastic objects is proposed in this paper. The proposed MOR scheme is capable of interpolating the aeroelastic objects in respect to the parameters. It attempts to reduce the number of degrees of freedom (DOF) from the pre-acquired high fidelity computational fluid dynamics (CFD) results. The number of DOF is reduced by implementing the proper orthogonal decomposition (POD) which converts the DOF in CFD nodes to those of POD modes and coefficients. Then the POD coefficients will be interpolated with respect to the parameters based on modified Nouveau variational autoencoder (mNVAE2). By mNVAE2, stable interpolation across the parameters will be conducted and accurately interpolated POD coefficients will be obtained. The interpolated aeroelastic objects are generated by multiplying parametrically interpolated POD coefficients in terms of the corresponding POD modes. The capability of the current MOR method for nonlinear aeroelastic objects will be demonstrated in this paper. It will be examined by interpolating the flow fields surrounding a stationary cylinder in terms of varying Reynolds number and prescribed plunging two-dimensional airfoil in terms of various plunging amplitudes. By those two examples, the current method is expected to accurately interpolate the flow fields of the parameterized aeroelastic objects efficiently.