Design with data-driven shape rule recognition through learning-based shape grammar interpreter

Shape grammars started from the concept of generating infitinte design alternatives with a limited number of shape rules through the shape itself. Shape grammar interpreter is a  computational framework dealing with the language of shapes using shape grammar formalism,  allowing users to transform one shape to another by applying shape rules. As a rule-based  method, it has been applied to various design problems, research, and education. Recently, the  Shape Machine has solved the long-standing sub-shape recognition problem in the shape  grammar interpreter in the 2D space. It enables users to explore designs precisely by visually  coding shape rules using vector data. This research offers a different perspective on coding the  shape rules from existing shape grammar interpreters such as Shape Machine, which focuses on  accurately recognizing and generating the shape. With the rapid growth of artificial intelligence  (AI) research, there have been increasing attempts to solve design problems using data-driven  methods like deep learning. This research proposes a learning-based shape grammar interpreter  by introducing deep learning, allowing designers to select shape data and train interpreters to  recognize shape rules. Designers understand and input the shape rules into the rule-based  interpreter based on visual representation. On the other hand, the learning-based interpreter  recognizes the shape rules by analyzing the relationships of elements selected by the designers  from the data. The designers choose the relationship that will be analyzed instead of analyzing it  themselves. This is determined by the designer's perspective on the design problem and may  offer different shape rules. After constructing the new interpreter, to understand designers'  experiences with it, a user study using simple geometric shape rules evaluated the learning-based  interpreter by comparing it with the existing rule-based shape grammar process. This research  demonstrates the potential of a learning-based interpreter to enrich the rule space through  different design experiences and unexpected design results, supporting current shape grammar  interpreters. However, it also shows the limitations of the design of the learning-based interpreter (challenging in data acquisition and processing, visualization, and understanding of rules) and  experiment (limited data selection, symmetry issues in comparative experiments).