Large Language Models demonstrate ability to generate multibody models

Researchers are investigating the capability of Large Language Models (LLMs) to generate multibody system dynamics models using natural language descriptions. The study explores whether these models can move beyond simple code replication to demonstrate a functional understanding of kinematics and dynamics.

Experimental findings show that LLMs, including those trained on the multibody code Exudyn, can successfully generate correct simulation models for simpler cases, often through zero-shot attempts. While the models demonstrate an ability to transfer knowledge into programming interfaces, complex scenarios frequently lead to modeling or programming errors. The research evaluated systems ranging from single mass oscillators to more complex setups involving mass points, constraints, and rigid bodies.