Accounting for Geometric Variability Effects on the Performance of Ducted Fan Blades Using Neural Networks and Multifidelity Optimization

Directly accounting for uncertainty and design variability in a multidisciplinary design process of turbomachinery blades is essential to ensure design robustness and reliability. However, performing such an optimization task can be computationally expensive, requiring many evaluations of the numerical model to compute the statistics of the blade performance. This paper proposes and implements an efficient multifidelity approach to robust optimization that leverages the benefits of neural networks, based on high-fidelity computational fluid dynamics (CFD) datasets and multiple inexpensive low-fidelity regression models for uncertainty quantification. The multi-information source is integrated using a multifidelity Monte Carlo method that optimally allocates the computational load based on relative evaluation cost and the strength of the correlation to achieve the relevant design statistics for the target objectives. The results showed that the proposed optimization strategy achieved mean aerodynamic performance with variations of 3.26%, and 16.67% for isentropic efficiency, and total pressure ratio respectively. Moreover, the total pressure ratio realized the highest reduction in standard deviation of 25%. By combining neural networks with cheap regression models the robust optimization solution was able to predict the statistical objective targets with reasonable accuracy.