The process of optimizing aerodynamic shapes with computational fluid dynamics is now a common practice. However, designing complex configurations presents significant challenges in both computational resources and the time needed to achieve results. Research in this field has focused on both intrusive methods, and on non-intrusive methods, like estimating the quantities of interest using surrogate models. Of particular interest in the latter approach are multi-fidelity methods, which, whether using gradient-based or gradient-free techniques, can significantly reduce the time needed to generate a design for complex aerodynamic shapes. This work extends an approach to aerodynamic shape design that uses a multi-fidelity hierarchical structure within an evolutionary optimization framework to 3D configurations. The various levels of fidelity can include both physical models of varying accuracy, solved numerically, and surrogate models based on supervised or unsupervised machine learning methods. These models incorporate information from physical models during both the training and generation phases. The core idea is to embed physics knowledge into the approximators through multi-fidelity modeling, whereby the approximators incorporate physical insights by using a low-fidelity model as input and yield the difference between the low- and high-fidelity models. The approach leverages a hierarchy of increasingly detailed solvers — from a potential flow with boundary layer correction to two different Euler and Reynolds Averaged Navier-Stokes solvers — to train a physically informed surrogate model. In particular, the open-source SU2 software is used to generate medium-resolution solutions for configurations around which an unstructured parametric grid can be easily created, while the in-house developed SIMBA Cartesian solver at CIRA is used to generate solutions on complex configurations automatically and easily, although at the cost of higher computational resource requirements. During the model training phase, this surrogate adjusts the loworder model response to approximate the high-fidelity solution. The goal of this approach is to minimize the size of the database required to effectively train an ML-based surrogate, balancing the need for extensive training data typical of machine learning tech techniques with the goal of minimizing the number of high-fidelity evaluations needed in the design process.