Résumé : In the present work, we introduce a multi-fidelity reduced-order model (MF-ROM) framework that effectively blends high-fidelity evaluations (accurate but expensive) with low-fidelity ones (approximate, less expensive) to predict the thermo-chemical state at unexplored operating conditions. The methodology combines Proper Orthogonal Decomposition (POD) for data compression, manifold alignment for blending information from high- and low-fidelity data, and CoKriging for regression. To assess the methodology, two-dimensional Reynolds-Averaged Navier Stokes (RANS) Computational Fluid Dynamics (CFD) simulations, along with a Chemical Reactor Network (CRN) derived from these simulations, are employed to develop the MF-ROM to predict the spatial fields of thermo-chemical variables at unexplored design conditions. Results show that the MF-ROM attains competitive predictive accuracy against the single-fidelity ROM built with only high-fidelity data for temperature and main chemical species distribution while considerably lowering the computational costs. This new framework allows to predict unexplored scenarios in a wide range of conditions, proving useful in preliminary design explorations, troubleshooting and addressing what if scenarios for a limited computational budget by incorporating simulations of different fidelities.