Tool-Augmented Agent for Closed-loop Optimization,Simulation,and Modeling Orchestration
Original reporting by arXiv (cs.AI)
In the world of industrial design, the journey from concept to finished product is an iterative dance between computer-aided design (CAD) and computer-aided engineering (CAE) simulations. Engineers meticulously craft designs, run simulations to test performance, then revise geometries based on feedback – a process often protracted and complex. The core challenge, termed the "CAD-CAE semantic gap," lies in efficiently translating complex simulation results into precise, valid geometric adjustments while juggling diverse design constraints. This labor-intensive loop has long limited the pace of innovation and optimal product development, demanding a more intelligent approach.
Bridging the gap
A novel framework, COSMO-Agent, offers a significant leap forward by empowering large language models (LLMs) to master this intricate closed-loop design process autonomously. Leveraging tool-augmented reinforcement learning, COSMO-Agent trains LLMs to orchestrate the entire workflow: from generating initial CAD and executing CAE simulations to parsing detailed results and iteratively revising parametric geometries. This interactive learning environment, reinforced by a sophisticated multi-constraint reward system, ensures design feasibility, toolchain robustness, and valid output until all specified criteria are satisfied.
Crucially, experiments reveal that COSMO-Agent enables even smaller open-source LLMs, when trained with its new industry-aligned dataset of executable CAD-CAE tasks, to substantially outperform much larger open-source and powerful closed-source models. This breakthrough marks a significant advance in the feasibility, efficiency, and stability of constraint-driven design, promising to accelerate product development and optimization across a wide array of industries.
COSMO-Agent marks a significant advance in the integration of artificial intelligence into critical industrial design processes. By effectively bridging the long-standing semantic gap between CAD and CAE, this novel framework empowers even smaller open-source language models to orchestrate complex design iterations with unprecedented autonomy and precision. The ability of LLMs to parse simulation feedback, interpret engineering constraints, and execute geometric revisions independently represents a fundamental shift from human-centric, iterative design toward highly automated, intelligence-driven optimization. This not only enhances design feasibility, efficiency, and stability, as demonstrated by its superior performance over larger and closed-source models, but also brings a new level of robustness to the entire development pipeline, promising to reduce errors and accelerate time-to-market.