GENERATIVE PRE-TRAINED TRANSFORMERS FOR BIOLOGICALLY INSPIRED DESIGN CONCEPT GENERATION

Abstract

For millions of years, biological systems in the natural world have changed to adapt to and 
endure their surroundings. Many of the characteristics they created might be useful and 
motivating for resolving technological issues in contemporary industry. This results in a 
particular kind of design-by-analogy known as bio-inspired design (BID). Despite the fact 
that BID has been shown to be a useful design tool, the gap between biology and engineering 
consistently prevents designers from using the method efficiently. In order to close the gap, 
we investigate the latest developments in artificial intelligence (AI) for a data-driven 
strategy. In order to automatically collect and map biological analogies and provide BID in 
the form of natural language, this work suggests a generative design technique based on the 
generative pre-trained language model (PLM). The basis PLM is the most recent generative 
pre-trained transformer, GPT-3. Depending on how flexible the issue space representation 
is, three different kinds of design idea generators are found and refined from the PLM. In 
order to evaluate the mapping relevance across the domains inside the created BID ideas, 
machine evaluators are also refined. During the conceptual design stage, the method is 
assessed and then used in a practical project to create lightweight flying automobiles. The 
results demonstrate that our method can provide BID ideas with high performance