Skip to content
COOLING OPTIMIZATION / NOTE 004

How AI optimizes data center cooling.

AI data center cooling optimization means adjusting cooling setpoints continuously to the actual heat load. This note explains what gets optimized, which methods exist and what to check before one of them touches a live system.

arflo Engineering · · 6 MIN

What cooling optimization actually changes.

A cooling system has a limited set of control inputs: fan speeds, pump speeds, valve positions, and the supply temperatures of air and water. Cooling optimization means choosing these values so that every component stays within its thermal limits while the cooling plant uses as little electricity as possible.

The term "AI" covers several different techniques for making that choice. They differ in what they need as input, how they behave in situations they have not seen, and how well an operator can follow their decisions. Those differences matter more in practice than the label.

Why fixed setpoints cost energy.

Cooling is usually commissioned for a demanding case: high IT load on a warm day, with a safety margin. A fixed setpoint then holds that operating point at all hours, including the many hours in which load or outdoor temperature are lower.

The cost of that margin is not linear. For fans and pumps, the affinity laws state that shaft power rises with roughly the cube of speed. Running a fan at 80% instead of 100% speed therefore needs about half the power in the ideal case. Real savings depend on the equipment, its efficiency curves and the rest of the system, which is why they have to be measured and not assumed.

Three ways to choose a setpoint.

Rule-based control uses feedback loops and schedules: if a temperature rises above its target, increase cooling. It is transparent and proven, but each loop looks at one measurement and reacts only after a deviation has occurred.

Data-driven control trains a statistical model on historical operating data and uses it to predict the effect of a setpoint change. The best-known example is the work DeepMind published with Google, which reported a 40% reduction in the energy used for cooling in the data centers where it was applied. This approach needs a large amount of representative data and is reliable mainly inside the range that data covers.

Physics-based control describes the system with energy balances: where heat is generated, where it is stored and how it is removed. Such a model can explain its predictions and has a defined behavior under conditions it was not tuned on, but it has to be built and calibrated for the system. Hybrid approaches combine the two: physics supplies the structure, and measured data adapts the parameters during operation.

Why AI workloads make the problem harder.

High-density GPU racks concentrate far more heat in one place than conventional servers, and many of them are liquid cooled. Their load can change within seconds when a training job starts or stops.

A room or return-air sensor registers such a change late and in attenuated form, because material and coolant store heat along the way. A controller that only sees room temperatures has to keep a larger safety margin. Using load and component telemetry lets a controller act on the cause instead of waiting for its delayed effect.

What to check before deployment.

Data: which measurements does the system actually provide, at what rate, and are timestamps and units consistent? Missing or stale values need to be detectable.

Boundary: what exactly is claimed to be saved, and how is it measured? A lower fan power does not automatically mean lower energy for the whole cooling plant. A fair comparison documents load, ambient conditions and thermal results for both cases.

Authority: does the software recommend setpoints to an operator, or does it write them itself? Direct actuation needs permitted ranges, defined responsibilities and a tested way back to the previous operating mode if data is lost or the model leaves its evaluated range.

Where arflo fits.

arflo follows the hybrid approach. It connects thermal physics with adaptive AI to model the path from chip to infrastructure and to consider load, local temperatures and cooling state together. How accurate such a model is, and what it saves, has to be established through validation on the specific system.