Modern engineering design workflows are inherently iterative, with each iteration typically requiring new sets of computationally expensive evaluations while previously generated data is discarded. This thesis proposes a methodology for reusing data from previous design iterations within a multi-fidelity (MF) Gaussian Process (GP) surrogate modelling framework. The data from a previous design iteration, referred to as the source, is fused with a limited number current design evaluations, referred to as the target, using Multi-Fidelity Co-Kriging (MFCK). The source data serves as the low-fidelity input and the target data as the high-fidelity input within the MFCK. The squared Pearson correlation coefficient is adopted as an a priori metric for assessing LF source quality, while a post-hoc discrepancy share diagnostic is introduced to evaluate the contribution of the LF source from the trained model’s internal perspective.
The methodology is validated on two structural design problems of increasing complexity: a three-dimensional cantilever beam problem with a readily available analytical solution and a six-dimensional stiffened cylindrical shell problem, evaluated using finite element analysis. The findings suggest that the effectiveness of the approach depends primarily on the quality of the source data: fusion of well-correlated designs can substantially reduce the number of required target evaluations while maintaining comparable precision and improved robustness. Poorly correlated designs, on the other hand, can degrade the performance below the singlesource counterpart.
... Read more