Do LLMs understand shapes? Exploring STL files for automatic CAD feature recognition

Abstract

The manufacture of industrial components requires a process planning stage where features such as slots, holes, and steps need to be identified in the Computer-Aided Design (CAD) models. Precise detection of these machining features is essential to generate accurate manufacturing instructions. Over the past four decades, automating this manual process has been an area of research known as Automatic Feature Recognition (AFR). To date, Convolutional Neural Networks (CNNs) are the state-of-the-art approach for this task. Nevertheless, given the increasing ability of Large Language Models (LLMs) to understand complex information, this work proposes employing LLMs for the AFR problem. In this study, ASCII-formatted STL files are treated as language-based data for a feature classification task. For doing so, Qwen, a pre-trained LLM with robust performance across diverse tasks, is compared to a CNN model that processes voxelised input. The results are validated to understand the capabilities and computational effort required by LLMs in the context of AFR, and to evaluate their understanding of new text-based formats such as STL. For this purpose, a dataset of 24 feature classes is used for a classification task. According to the results, LLM-based methods demonstrate an understanding of STL data, revealing potential for feature classification in this field.