MechRAG: a multimodal large language model for mechanical engineering

Abstract

Engineering design and manufacture are inherently multimodal activities in which engineers consult and produce diverse data and representations across various engineering disciplines and product lifecycle stages. Although well-established digital formats exist for these representations, their use remains restricted within specialist applications, creating silos that limit cross-domain integration. Here we introduce mechanical retrieval-augmented generation (MechRAG), a multimodal large language model architecture designed to unify information from multiple engineering representations typically found in computer-aided engineering and computer-aided design environments. Results demonstrate that MechRAG achieves high accuracy in routinely performed mechanical activities such as data-management or classification tasks, and effectively replicates engineer-level reasoning in more inferential and subjective contexts. Our findings suggest that such conversational interfaces enhance engineering productivity, facilitate more interactive paradigms, and drive transformative workflows across various stages of design and manufacturing.

Shuang Li and colleague propose a multimodal, retrieval-augmented, large language model MechRAG. It integrates heterogeneous CAD/CAE digital assets into its responses to engineering questions delivered as prompts in a conversational interface