Incremental sampling methods for multi-fidelity surrogate modelling of a MILD combustion furnace

This work presents a framework for the generation of a multi-fidelity reduced order model (MF-ROM) of a combustion furnace operating in MILD conditions. The framework combines data compression by Proper Orthogonal Decomposition, information transfer between different fidelity models by Procrustes manifold alignment, and interpolation by CoKriging. Two- and three-dimensional simulations of the MILD combustion furnace with design parameters of air injector diameter, fuel composition, and equivalence ratio are fused to build the model. Moreover, a typical question related to MF-ROMs is tackled, that is, how many high-fidelity simulations should be employed to get a favorable trade-off between accuracy and training cost. Incremental sampling strategies are implemented to answer this question, and the results show that approximately half the training cost can be saved to obtain the same error values.