Ride comfort of high-speed elevator systems is often influenced by complex confined aerodynamic effects, including unsteady pressure fluctuations, elevated aerodynamic drag, and flow-induced noise [1]. However, accurate aerodynamic optimization of elevator cabins is challenging because of the enormous computational cost in modeling the high-fidelity coupling between geometric design variables, dynamic factors and flow physics. To address these challenges, this study proposes a physics-informed multi-fidelity optimization framework that integrates CFD-based simulations with physics-informed neural networks (PINNs) for constrained aerodynamic optimization. In the proposed framework, low-fidelity models are constructed using simplified Reynolds-averaged Navier–Stokes (RANS) formulations and flow representations, while medium transient CFD simulations are selectively employed for critical flow regimes characterized by strong confinement and unsteady pressure oscillations [2-3] and HF results are conducted by physical experiments. A PINN-based surrogate model is introduced to embed the governing Navier–Stokes equations and boundary constraints into the learning process, which serves as a physics-regularized bridge between different fidelity levels. An enhanced constrained PSO-ADAM optimizer is then integrated to adaptively refine the MF model and search the global best point, in which PINN generates fitness function for the PSO and provides gradient information for ADAM optimizer. With the cabin’s geometric configuration fixed, this study optimizes the blockage ratio, door-gap dimensions, and bracket spacing to mitigate aerodynamic effects and optimize passenger ride comfort during high-speed operation. The operational velocity, maximum pressure fluctuation rate, and both horizontal and vertical vibration accelerations are considered as optimization constraints. Numerical results demonstrate that the proposed framework achieves significant reductions in CFD evaluations while maintaining high predictive accuracy. Comparative studies confirm that the proposed physics-informed multi-fidelity optimization framework outperforms conventional single-fidelity and data-driven multi-fidelity surrogate approaches including Expectation Improvement per-unit cost (EIpuc) and cost-aware lower confidence bound (cLCB) based co-Kriging methods in terms of computational efficiency and optimization accuracy.