Exploring Latent System Design Description and Requirement Similarities Using Transformer Based Contextual Embeddings

Abstract Engineering design is heavily influenced by system requirements, as they span the entire design process. Generally, requirements are not utilized to convey system information outside the design process. System information is generally conveyed to stakeholders, such as end users, through generalized descriptions. While requirements and system descriptions both represent the same system to different audiences, the understanding of how each class is related is not well explored, which Large Language Models have now allowed us to do. We employ a method of using pre-trained contextual embeddings from RoBERTa and the resulting UMAP projections to draw conclusions about local and global embedding space characteristics from text-based system information. Though some systems may exhibit overlap or regionalization, results demonstrate that each class (requirements and descriptions) can be utilized to compare similarities between systems from a functional perspective but cannot be directly compared when trying to use requirements and descriptions jointly in the embedding space. The degree to which a system is novel is distinct in the description-based embedding space only by means of the underlying technological principles and not from the form. The implications of this have impacts on designers attempting to leverage LLMs in an attempt to understand how products are differentiable and novel in comparison to other existing systems when either the function of the product is unique, or the physical aspects of the product are significantly different.