Publications

Effective access to aerodynamic flow fields reveals essential physical insights (such as shock waves and boundary layer separation) related to aerodynamic efficiency. Thus, an adequate description on flows fields is advantageous for airfoil design optimization with reference to achieving desirable aerodynamic performance. Conventional flow field computation, however, relies on high-fidelity but time-consuming computational fluid dynamics models. Alternatively, surrogate modeling (such as reduced-order modeling) arises for fast flow field prediction, but randomly sampling an input space could result in intensive training data cost. To tackle these issues, in this work we propose a fully automated adaptive sampling strategy for the proper orthogonal decomposition-based reduced order modeling. In particular, we leverage a mathematical sampling criteria (known as potential) for intelligent selection among sample candidates. Moreover, we propose to select the number of reduced-order bases and switch to a second strategy on the fly, instead of user-defined and fixed throughout the whole process. Therefore, "fully automated" refers to the fact that the sampling strategy exerts fully automated control over the adaptive selection to the samples and bases. We demonstrated the proposed sampling strategy on airfoil flow field prediction within transonic regime in this work. Results showed that the surrogate predictions based on random sampling via Latin hypercube sampling had a error of 1.6E−3 over the input space of Mach number ([0.55, 0.75]) and angle of attack ([-1, 3] deg). In contrast, the fully automated strategy made a mean error of 7.69E−4 and a small standard deviation of 5.72E−5 over the input space for 10 different trials. The worst predictive performance has an error of 8.01E−4. Furthermore, we compared the fully automated strategy against the fix-switch strategy. The mean error of fixed-switch strategy varies from 6.83E−4 to 8.01E−4 depending on when to switch to the second strategy. The worst predictive performance has an error up to 1.4E−3. Note that we set 40 training samples as the total computational budget and compared predictive performance under this condition. Therefore, we can conclude the fully automated algorithm has promising predictive performance with sufficient robustness with respect to sample candidate pools. Note that the proposed strategy is also extensible to other engineering fields.