UT deploys machine learning to boost airflow sensor performance

Researchers from the University of Twente (UT) have used machine learning to improve “a simple and straightforward sensor” for measuring air speed, while simultaneously increasing its functionality. The accuracy of the device improved by over 500 percent compared to existing methods. These findings could help optimize conditions in vertical farms and manage airflow in data centers.

“With just one flow sensor, we can extract far more information. That means we can use less expensive hardware and invest more in advanced analysis. Our sensor, as a research prototype, isn’t perfect in the conventional sense – it’s more sensitive to those external factors. Instead of only capturing flow data, it also picks up temperature, humidity and other environmental clues. Unlike the wider sensors community, we see this as an opportunity, not a hindrance. It enables us to extract more information using one sensor,” says Thomas Leigh Hackett, PhD candidate at the UT’s Integrated Devices and Systems department.

The MEMS device is called a thermal anemometer, which measures airspeed through heat transfer from a heating element into the flowing medium. By applying machine learning to different signals from every individual sensing element, the angle, airspeed and relative humidity can be independently extracted, even though the design is quite simple.

Furthermore, the miniature sensor (0.16 cm²) was fabricated with a straightforward silicon-on-insulator (SOI) fabrication procedure, integrating both the sensing and computing elements. “This way, we’re cutting production costs in half because we don’t need the MEMS foundry: everything is done in the same facility,” Hackett explains.