AI engineer tests 1,460 solutions in 17 hours, data center electricity bills expected to save $227,000 annually.

Beating AI News Flash: YC-incubated startup Standard Physics has launched AI engineer Archimedes. It can design engineering solutions on its own, run physics simulations, and continuously revise based on results to find better-performing designs.

In a data center in the United States, cooling equipment was already running at full speed, yet the hottest server rack still reached 35.6°C. Through simulation, the AI found that some cold air was not being delivered to the right locations, and hot air exhausted by servers was flowing back into the racks.

The AI first tried adjusting air supply vents and isolating hot air, then optimized the operation of the cooling equipment. The energy-saving plan found in the first round looked very effective, but it ignored the actual capabilities of older equipment. After review found it could not be implemented, the AI added equipment constraints and searched for a new plan.

The entire process lasted 17 hours, with a total of 1,460 designs tested. In the end, two retrofit plans were selected. According to the company's simulation estimates, a full retrofit could reduce the data center's total electricity consumption by 10.9%, bring the hottest rack down to 29.2°C, and save about $227,000 in electricity costs per year. The other low-cost plan requires less than one-quarter of the retrofit cost of the full retrofit while achieving 79% of its energy savings.

The AI also tested 1,333 turbine blade designs through 2D simulation within 30 hours. While maintaining the same airflow turning requirements, the final design reduced energy loss by 25.5% compared with the original blade, surpassing an optimized design published by a Belgian research team in 2018.

The entire system is completed through the collaboration of two parts: Archimedes is responsible for proposing designs, comparing results, and deciding how to revise next; Reynolds is responsible for running simulations. It can convert problems described by engineers in natural language into OpenFOAM simulation tasks, automatically build models, generate computational meshes, run calculations, and analyze results.

Standard Physics has open-sourced Reynolds, which is responsible for running physics simulations, and anyone can use and modify the code for free. The complete AI engineer Archimedes, which can autonomously design solutions, repeatedly test them, and find optimal solutions, has not yet been open-sourced and requires contacting the official team for use.