Search papers, labs, and topics across Lattice.
This paper introduces a framework that leverages LLMs with function calling and agent workflows to automate CAD geometry generation from textual prompts. The framework explores different agent workflows and LLMs, finding that workflows incorporating automated visual feedback, particularly with multimodal LLMs like ChatGPT-4o, perform best. A case study demonstrates the framework's application in topology optimization and additive manufacturing, highlighting its potential to reduce manual effort in design processes.
Automating CAD design from text prompts is now feasible, with visual feedback loops boosting performance, especially for multimodal LLMs.
ABSTRACT: Design generation using traditional Computer-Aided Design (CAD) tools remains a labor-intensive and manual task. This paper introduces a framework for automating CAD geometry generation using Large Language Models (LLMs) with function calling and agent workflows. The framework enables both expert and novice designers to use textual prompts to automatically generate CAD code. We evaluate it with five LLMs and four agent workflows. The agent workflow incorporating automated visual feedback outperforms the others, especially with multimodal LLMs like ChatGPT-4o. A case study shows its use in topology optimization and additive manufacturing with minimal human input. Remaining challenges include limitations in spatial reasoning, prompt dependency, and workflow adaptability. Future work should focus on improving design-for-manufacturing capabilities, visual tools, and evaluation benchmarking.
ETH