A K-Medoids-Based Shape Clustering Method and Its Applications in Generative Design and Optimization Systems

Abstract

As the number of design candidates in generative and design optimization systems is often excessive and overwhelming, with similar and redundant shapes of design candidates evolving, there is a need for an articulation mechanism that assists designers in the exploration and examination of the design set in a feasible manner. This work is focused on introducing a new Generative Design System (GDS) that facilitates the designers’ interaction with such systems and accommodates decision making in the process. The proposed system incorporates an innovative Shape Clustering using K-Medoids (SC-KM) method into other routine processes of parametric modeling and design optimization.

The research methods include an extensive literature study, experimenting, prototyping, and validation procedures. A prototype was carried out, as the apparatus to demonstrate and test the proposed GDS, and the clustering method. In developing, demonstrating, and testing the prototype, three test-cases were pursued. Within the prototype, at first, a process of parametric form generation was carried out to initiate a design model parametrically to allow for a possibly heterogeneous or/and similar set of design options to be produced, in an algorithmic manner. Second, a design optimization process was pursued where the initial parametric model was subjected to building performance evaluation inside a Multi-Objective Evolutionary Algorithm, such as in Test-case 3. In the third process, the new SC-KM method was formulated and applied, using two functionalities: (1) a grid-based shape descriptor was used for a pair-wise shape comparison with the implementation of the Hungarian’s algorithm, carried out to find the shape difference score matrix for the analyzed shapes, (2) K-Medoids clustering was employed to group the design shapes into different subsets, each of similar shapes, and identify each group’s Medoid–the representative shape in the group. Applying the algorithmic definition to the samples of three test-cases, the results of the SC-KM showed satisfactory clusterings. Furthermore, verification procedures were conducted for each test-case, and in particular, external validation with the calculation of clustering evaluation metrics was pursued in Test-case 2.

In contrast to the accepted practice of current generative design systems that lack organizational methods, the significance of this work is to expand and advance such systems incorporating cluster analysis as a big-data management strategy and a potential solution. The research provides contributions through the following: (1) introduction and illustration of a fully working prototype of a new generative design system, (2) development, testing, and validation of a new package of algorithms for the developed SC-KM, method. The package of plugins and algorithms will be made available for designers to download as an open-source in a visual programming interface, to be applied to a wide range of related design problems.