Accurate numerical prediction of turbomachinery-relevant internal flows, including in highly curved ducts, remains a significant challenge for Reynolds-Averaged Navier–Stokes (RANS) methods. These flows are characterized by strong streamline curvature, secondary motions, flow separation, and large-scale unsteadiness, all of which are difficult to capture using conventional eddy-viscosity closures. While scale-resolving simulation (SRS) approaches, such as Large Eddy Simulation (LES) and hybrid RANS–LES methods, can reproduce these features more faithfully, their computational cost renders them impractical for parametric studies or design optimisation. This creates a gap between affordable low-fidelity models and highly accurate but prohibitively expensive high-fidelity simulations. This study investigates the potential of multi-fidelity modelling (MFM) based on Gaussian Process Regression (GPR) as a means of bridging this gap. The Military Engine Inlet Research Duct (MEIRD), a highly serpentine intake geometry developed at the University of the Bundeswehr Munich, is selected as a test case. This configuration has been studied both experimentally and numerically [1], and previous studies have highlighted its complex three-dimensional flow physics, making it a stringent test for advanced turbulence modelling strategies. A hierarchy of numerical simulations will be conducted including steady and unsteady RANS simulations employing advanced v2f-type closures as well as hybrid temporal LES [2] approaches derived from these formulations. The predictive capabilities and limitations of each modelling level will be systematically assessed. The resulting datasets will then be combined using GPR-based MFM to construct accurate, low-cost surrogate models for selected quantities of interest. To enable efficient handling of large, correlated flow datasets, dimensionality reduction techniques based on Proper Orthogonal Decomposition (POD) will be employed in conjunction with GPR [3]. Time permitting, uncertainty quantification will be incorporated to assess sensitivity to variations in operating condition. The proposed framework aims to enable reliable, physics-informed predictions of complex internal flows at a fraction of the computational cost of full SRS over the parameter space, with direct implications for preliminary design, optimisation, and robust engineering analysis.