<p>The realm of high-end equipment necessitates the development of plate-type metastructures integrating lightweight, high static load-bearing, and superior low-frequency dynamic performance. However, the rapid on-demand inverse design of such plate-type metastructures remains challenging. This study proposes a “data + non-data” hybrid-driven inverse design (HDID) method for a class of locally resonant plate-type multifunctional metastructure (LRPMM). The LRPMM primarily comprises two spiral plates, a square honeycomb core, and a cylinder. Its load-bearing and dynamic performance are characterized by bending stiffness and flexural wave bandgap, respectively, both investigated through theoretical and numerical methods. The bandgap is predicted through the synthesis of a discrete dimensionality-reduction strategy and the plane-wave expansion method. Based on this, a dataset is generated, and a forward prediction network model is trained. Then, a “data + non-data” model is constructed to target the customization of bending stiffness and flexural wave bandgap. The core of the hybrid-driven tandem neural network model lies in the fact that pre-trained forward prediction network models are employed for predicting the flexural wave bandgap with a strongly nonlinear property, while physical equations are used to characterize the bending stiffness with a weakly nonlinear property. Results demonstrate that the proposed HDID method can improve the design efficiency by over 10% while ensuring high accuracy, compared with the conventional data-driven inverse design (DDID) method. Notably, the proposed method enables the design of the metastructure with a bending stiffness of 18315 N m, an equivalent density of 2 g cm⁻³, and effective flexural wave suppression in the frequency range of 55.9–94.5 Hz. Finally, vibration suppression experiments are conducted to validate the effectiveness of the proposed HDID method.</p>