Practical aerodynamic analysis involves the numerical solution of complex flows involving compressibility, discontinuities such as shock waves, and multi-physics phenomena. Advanced numerical schemes and high-performance computing are necessary to resolve RANS or LES calculations. These factors make aerodynamic design optimization more challenging due to the high computational cost in terms of time and memory. Traditional surrogate models approximate only a limited number of performance metrics, which represent integrated quantities derived from flow state variables computed by CFD analysis. As a result, detailed flow features in regions of interest cannot be investigated during the design process when surrogate-based analysis and design approaches are used. In the present study, various AI-based methods are introduced to predict detailed flow state variables in the domain of interest at a fraction of the computational cost required for full-order CFD calculations. Conditional U-Net, CNN-LSTM, and PointCloud U-Net methods, as well as a traditional POD-DEIM method with grid adaptation, are effectively utilized to provide high-fidelity CFD solutions across a wide range of flow regimes, from incompressible to hypersonic flows. Physics-Informed Neural Networks (PINNs) and neural operator networks are also investigated as neural PDE solvers for compressible flows. Issues related to computational grids, flow discontinuities, and solid boundary detection are discussed. In summary, these AI-based reduced-order models enable a powerful digital twin framework with an integrated physics engine.