OpenFOAMGPT: The world’s first AI engineer

By combining a Large Language Model (LLM) that can generate text with software for numerical fluid mechanics on a single platform, researchers have created an AI-powered engineer that can automatically solve problems. It might sound unbelievable. But it’s true. In a joint project, a group of scientists from the SimTech Cluster of Excellence, the Faculty of Aerospace Engineering, and Geodesy have created the world’s first AI-powered engineer that works quickly and reliably.

Four agents share the tasks of the AI engineer. These include pre-processing, prompt creation, simulation, and post-processing. The lead agent can understand and analyze requests from LLMs. For example, it can automatically identify variables that are required to solve the task and understand whether the query involves simple or complex geometries. The prompt creator constructs simulation instructions that are precisely tailored to the task at hand.

At the heart of the system is the free software package OpenFOAM, which carries out the simulations. OpenFOAMGPT automatically configures simulations based on prompts and task requirements. The simulation is started automatically. If it fails because of a faulty configuration, the error log is processed in a structured manner and returned to the LLM together with the original prompt. This creates a feedback process in which configuration files are intelligently improved until a successful simulation is achieved. The rework agent analyzes the simulation results and creates comparison diagrams and charts.

The design of a multi-agent framework

To test the reliability of the AI engineer, the scientists carried out various experiments. They selected five case studies covering a wide range of computational fluid dynamics tasks. Examples ranged from simulating the flow of a viscous liquid through a straight channel under pressure to modeling multiphase flow in porous media (such as in oil drainage processes) to analyzing turbulent airflow around a motorcycle at varying speeds.

Because reproducibility is an indicator of reliability, some of the simulations were carried out up to a hundred times, and the result was the same each time. “That surprised us—and also scared us a little,” says Xu Chu. “For example, if I upload a picture of myself to ChatGPT and ask 10 times whether I look good in it, I get 10 different answers. Of course, that’s not possible with engineering issues,” he says. “My doctoral researchers and I weren’t able to sleep well some nights because we saw what it means when the system works so reliably.”

According to Xu Chu, this could revolutionize the field and make intelligence widely accessible. He doesn’t even want to imagine what that would mean for his colleagues and engineers. “We have found our own replacement,” he says. Just as Xu Chu and his team taught the LLM to use OpenFOAM and do its job, the same could be applied to fields other than fluid mechanics. 

OpenFOAMGPT analyzes and creates charts

Xu Chu has taken it a step further by integrating additional agents that read scientific books and publications, develop research ideas, analyze OpenFOAMGPT results, and write scientific manuscripts. He calls the system Turbulence.ai, and it is the world’s first AI scientist in the field of fluid mechanics. Turbulence.ai conducts completely autonomous research and publishes high-quality specialist articles.

It poses and formulates hypotheses, plans simulations and checks them for novelty and feasibility, corrects errors, prepares the results visually, and finally writes a complete scientific article without the need for human intervention. The first manuscript on two-phase displacement processes in porous media is already finished.

Xu Chu is still unsure whether it should be submitted for peer review. “Because I have already published a lot myself and also review manuscripts as a reviewer, I can say that it is not top quality. But it is good enough to submit and publish in an above-average journal,” explains Xu Chu. Because many questions in fluid mechanics remain unanswered, the AI scientist could make a real contribution to the field. Xu Chu sees himself as a source of ideas and aims to inspire researchers in fluid mechanics and other disciplines and open up new possibilities for them. He is also open to collaboration with industry.

Xu Chu would like to thank team members Jingsen Feng, Ran Xu, Yupeng Qi, Sandeep Pandey, and Wenkang Wang.