AI Engineer Develops 1,460 Solutions in 17 Hours, Data Center Expected to Save $227,000 Annually on Electricity Costs

It can autonomously design engineering solutions, run physical simulations, and continuously modify based on results to find better designs.

In a data center in the United States, cooling equipment is running at full speed, yet the hottest server cabinets still reach 35.6°C. AI discovered through simulations that some cold air was not reaching the correct locations, and the hot air expelled by the servers was flowing back into the cabinets.

AI first attempted to adjust the air supply vents and isolate the hot air, then optimized the operation of the cooling equipment. The first round of energy-saving solutions seemed effective but overlooked the actual capabilities of the old equipment. Upon review, it was found to be unimplementable, so AI added equipment constraints and searched for new solutions.

The entire process lasted 17 hours, during which 1,460 designs were attempted. Ultimately, two renovation plans were filtered out. According to the company's simulations, a comprehensive renovation could reduce the total electricity consumption of the server room by 10.9%, lowering the temperature of the hottest cabinet to 29.2°C, saving approximately $227,000 in electricity costs annually. Another low-cost plan could achieve 79% of the energy-saving effect of a full renovation at less than a quarter of the renovation cost.

AI also tested 1,333 designs for turbine blades through two-dimensional simulations within 30 hours. The final plan reduced energy loss by 25.5% compared to the original blades while maintaining the airflow direction requirements, surpassing an optimization design published by a Belgian research team in 2018.

The entire system is completed by two parts working together: one is responsible for proposing designs, comparing results, and deciding the next steps; the other runs the simulations. It can convert problems described in natural language by engineers into OpenFOAM simulation tasks, automatically modeling, dividing computational grids, running calculations, and analyzing results.

The part responsible for running physical simulations has been open-sourced, allowing anyone to use and modify the code for free. However, the complete AI engineer capable of autonomously designing solutions, repeatedly testing, and seeking optimal solutions has not been open-sourced and requires contacting the official team for use.