Shape optimization plays a crucial role in improving the performance of thermal–fluid devices, and there exist strong industrial demands for their optimal designs. The adjoint method has been extensively studied and successfully applied to steady flow problems, while their extension to unsteady flows remains challenging. This difficulty arises from the complex derivation of adjoint equations and the high computational cost associated with time-dependent adjoint analysis. Physics-Informed Neural Networks (PINNs) [1] have recently emerged as a promising alternative, as they enable the evaluation of sensitivities with respect to design variables through automatic differentiation without requiring adjoint formulations. Meanwhile, in the conventional PINNs, wall boundary conditions are often imposed as soft constraints in the loss function. As a result, the fluid velocity on a solid surface may not completey vanish. In addition, for unsteady flow problems, the conventional PINNs often underrepresent a time-derivative term and therefore the solutions tend to converge to steady ones [2]. To overcome these issues, Data-Assisted PINNs have been proposed, in which experimental data or CFD data are incorporated into the training process [3]. However, most existing approaches rely on high-resolution and/or high-fidelity data, thereby limiting their practical applicability. In this study, we propose a Hard-Constraint Data-Assisted PINN (HCDA-PINN), in which the no-slip condition is enforced as a hard constraint, while undersolved CFD data are additionally incorporated to assist to reproduce the unsteadiness of the flows. The HCDA-PINN is applied to unsteady laminar flow around a 2D cylinder at Re=100. The results demonstrate that the Kármán vortex street is accurately reconstructed both qualitatively and quantitatively. Furthermore, the proposed framework successfully enables shape optimization in an unsteady flow field.