Bridging the CAD-CAE Semantic Gap
Industrial design is frequently bottlenecked by the disconnect between Computer-Aided Design (CAD) and Computer-Aided Engineering (CAE). Translating simulation feedback into actionable geometric modifications requires deep domain expertise and the ability to navigate complex, coupled constraints. COSMO-Agent (Closed-loop Optimization, Simulation, and Modeling Orchestration) addresses this by treating the entire design-simulation-revision process as an interactive reinforcement learning (RL) environment.
The COSMO-Agent Architecture
The framework teaches LLMs to act as orchestrators for external engineering tools. The process follows a closed-loop cycle:
- CAD Generation: The agent generates parametric geometries.
- CAE Solving: The agent triggers simulation tools to test the design.
- Result Parsing: The agent interprets simulation feedback.
- Geometry Revision: The agent iteratively modifies the design based on feedback until all constraints are satisfied.
To ensure stability and industrial utility, the authors implemented a multi-constraint reward function. This function optimizes for three distinct pillars: design feasibility, the robustness of the toolchain integration, and the validity of the structured outputs. By training on a new, industry-aligned dataset covering 25 component categories, the model learns to navigate the specific, rigid requirements of engineering workflows rather than just generating generic text.
Performance and Practical Impact
Experimental results demonstrate that training smaller, open-source LLMs with the COSMO-Agent framework allows them to outperform both larger open-source models and strong closed-source models in constraint-driven design tasks. The framework significantly improves efficiency and stability, proving that specialized RL fine-tuning is more effective for technical orchestration than relying on the general reasoning capabilities of larger, unspecialized models.