Synthetic Dataset of APC 14x13 SP Propeller
Description
The strong linear relationship can limit the generalizability of GP models in complex scenarios. In particular, if the models are trained on data with high correlation, the model's capability to capture nonlinear discrepancies such as stall onset or wake interactions remains untested. Often, the design space of the drone propeller is highly complex, and the associated experimental data and various numerical modeling approaches are noisy in nature due to measurement uncertainties, experimentation limitations, and the simplified physical assumptions of the numerical models. Thus, to investigate such a scenario, we designed a synthetic low-fidelity function that introduces bias and loss of fidelity with an objective to serve as a controlled test case to analyze nonlinear biased fidelity mappings through the deliberate deceptive function as referred to Forrester et al. The synthetic function was developed specifically for the APC 14×13 SP propeller. The propeller was selected as an ideal model due to its nature of highlighting nonlinear aerodynamic effects and potential fidelity mismatch. By focusing on one propeller, the analysis avoids redundancy while providing a challenging test case that is adequate to assess the models' robustness. The true experimental data (high-fidelity) for the chosen propeller can be found in the literature (Vizeu et al.) along with the numerical modeling. The design introduces nonlinear distortions such as multiplicative sinusoidal scaling and noise, combined with high fidelity to mimic a realistic yet challenging fidelity relation. This setup allows us to test the GP models rigorously and determine whether nonlinear MF models such as NARGP provide better performance under such a setup and linear models like AR 1 and MFK degrade in accuracy and generalization. To replicate a complex nonlinear setup, two low-fidelity synthetic functions are designed for APC 14x13 SP, one each for Ct and Cp. The designed functions reflect the high-fidelity trends using independent scaling and residual terms and it follows the foundations of Perdikaris et al.