On the effectiveness of Large Language Models in the mechanical design domain

One widely used method to document and communicate design decisions is with Computer Aided Design (CAD) software. CAD software allows engineers and designers to create, modify, analyze, and optimize their design, while also documenting various aspects of the design, allowing them to communicate their design choices to others. Using CAD, designers create assemblies of parts, and often use natural language to name each individual part, as well as the assembly itself, for documentation and collaboration purposes, as shown in Figure 1. Often, domain-specific language is used to label parts, which raises questions about the effectiveness of current natural language processing (NLP) techniques in understanding this domain-specific mechanical design language. NLP techniques have been successfully used to understand the complexity of text and language for multiple tasks such as translation, question answering, and summarization. For these tasks, techniques such as word embed-dings, transformers, and attention help navigate the unorganized structure of language. Within NLP techniques, attention-based algorithms have shown state-of-the-art results on NLP tasks as text classification outperforming most featured-based representations methods, e.g., word2vec, Glove, CoVE en ELMo [1]. In this paper, we will explore the use of these techniques in the mechanical engineering domain. While some literature has leveraged NLP techniques for data-driven design, little work has been done specifically on natural language used to label CAD assemblies.