Design Space Reduction using Clustering in Aircraft Engine Design

Abstract

During the initial stages of design, it is critical to conduct design space exploration studies and evaluate the design solutions against given criteria. These criteria correspond to the goals that the current design needs to achieve. When there is high dimensionality, the design space becomes very large which is hard to evaluate comprehensively. Hence, instead of searching the entire design space, multiple clusters are obtained based on the design performance objectives. By only exploring the subsets of the design space, the computational cost is diminished, making the design space exploration studies faster. As a use case, conceptual design of turbofan engines is illustrated since propulsion design necessitates a high dimensional design input. Initial sampling is performed using the design of experiments approach, then the design space is clustered based on specific fuel consumption at cruise and engine weight. Both of the design performance parameters are to be minimized to achieve the optimal design point. The best cluster is chosen and regarded as the new space to be explored. Same procedure is applied by following an iterative design space reduction methodology. Two different clustering algorithms are employed and their performance assessment is made. It is found that K-means clustering11K-means clustering is an unsupervised learning technique, which assigns each data point to a group, given the attributes. K represents the number of groups found in the data. approach reduced the design space more than the Gaussian Mixture Model clustering technique22Clustering of data using Expectation-Maximization (EM) algorithm for Gaussian mixtures.. Additionally, the truth model33The correct model that is used for verification purposes. is obtained using the Pareto Frontier44Pareto Frontier is the set of design points optimizing the design problem. of the design dataset. The best design point in the final design space is compared against the truth model. It is found that the accuracy is higher when the k-means clustering approach is used.