Abstract
This paper describes a modified variable complexity modeling (MVCM) framework that uses a neural network to replace the Taylor series after several warm-up iterations. The MVCM framework with an additive scaling function is most efficient in terms of high-fidelity function evaluation savings compared with traditional variable complexity modeling (VCM) among multiplicative and hybrid scaling functions. The MVCM framework achieves 59.1 and 68.6 % savings in high-fidelity function evaluations for one-dimensionally and two-dimensionally constrained problems, respectively, compared with the VCM method. The MVCM framework provides a larger trust region than the VCM due to the global behavior of neural networks. The MVCM solution also converges closely to the high-fidelity function. The MVCM framework is integrated with an in-house low-fidelity aircraft design synthesis program and a high-fidelity analysis (AADL3D) for the conceptual design of multidisciplinary regional jet aircraft (RJA) to enhance the optimal RJA configuration compared with low-fidelity analysis results. The optimal RJA wing configuration using the MVCM framework provides more realistic and reasonable configurations compared to the results of low-fidelity analysis with a short turnaround time.
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have