Parametric Design Optimization in Computational Aerodynamics using Artificial Deep Neural Networks
Chintan Patel, Chad Iverson · AIAA Propulsion and Energy 2020 Forum · 2020
Application of machine learning based surrogate model and its utility in evaluating very large design space is demonstrated using meta-modelling based design optimization (MBDO). Carefully trained artificial neural networks (ANN) are shown to provide highly accurate surrogate models for CFD design optimization that is capable of producing globally optimal solution. These ANN surrogate models enable evaluation of large designs space consisting billions of designs with high accuracy, both rapidly and inexpensively. Diffuser airfoil shape of a centrifugal compressor assembly was optimized for maximum isentropic efficiency using CFD analysis for design evaluation and using ANN as a surrogate model for optimization. Efficacy of the approach is then demonstrated by comparison with very popular stochastic optimization class of method – genetic algorithms (GA). Present approach significantly reduces design time when compared to GA. GA (and a vast array of stochastic optimization methods) frequently get stuck in local optima, and global optima is illusive. Since entire design space can be evaluated with the presented approach, global optima can be found relatively inexpensively. It is shown that with same number of design evaluations, ANN finds a vastly superior compressor design compared to GA. For simplicity, single objective optimization is demonstrated, but the approach can seamlessly extend to reduced order modelling, multi-objective optimizations, multi-disciplinary design optimization, and uncertainty quantification.