UMD Develops ML Tool to Predict Uneven Airflow in AC Systems

Graduate research assistant Brian O'Malley with the wind tunnel used for in-house testing to validate the machine learning framework.

A University of Maryland team has developed a machine learning (ML) framework that could boost air-conditioning efficiency by quickly predicting changes in airflow velocity that affect heat exchanger performance. Researchers from UMD’s Center for Environmental Energy Engineering presented the work in the September 2026 issue of International Journal of Refrigeration.

Uneven airflow — known as airflow maldistribution — is a major cause of performance degradation in air-to-refrigerant heat exchangers. This issue is common to central air-conditioning systems, where compact ducts motivate heat exchanger designs that are prone to uneven airflow. Some studies indicate that airflow maldistribution can cause up to 65% reduction in cooling capacity. To compensate, heat exchangers may be oversized, increasing material costs, system footprint and refrigerant charge.

“If we can better predict airflow, the heat exchangers in these systems won’t have to be so oversized,” says the paper’s first author, Brian O’Malley, a CEEE graduate research assistant. “Our ML framework is an enabling technology that can lead to further innovation in heat exchanger design. By providing fast and accurate predictions, it gives engineers more information to make better design decisions. Ultimately, that could mean more affordable, higher-efficiency systems for consumers.”

Co-authors are post-doctoral researcher James Tancabel and CEEE Director and Research Professor Vikrant C. Aute, both with the UMD Department of Mechanical Engineering.

The team’s ML models were trained on results generated by porous media computational fluid dynamics (CFD) simulations. The resulting ML framework generates predictions at least 100,000 times faster than full porous media CFD.

"By providing fast and accurate predictions, it [the ML framework] gives engineers more information to make better design decisions. Ultimately, that could mean more affordable, higher-efficiency systems for consumers.”

CEEE graduate research assistant Brian O'Malley

“The ML framework offers an additional tool,” O’Malley says. “Early in the design cycle, you want to evaluate many options. This tool lets you do that without building a test setup or running a full simulation. It’s quick and easy to use.”

The team validated the machine-learning framework using three sets of test data. Two datasets came from previously published studies: one from an A-shaped residential air-conditioning fin coil and another from a U-shaped outdoor unit. The third came from in-house testing at the Daikin Energy Innovation Laboratory, using a prototype heat exchanger section designed by UMD-CEEE researchers.

For every case, the framework was able to rapidly and accurately predict the airflow maldistribution profile at the inlet of the heat exchanger, showcasing its general applicability to many different heat exchanger types of interest.

For O’Malley, a doctoral candidate in mechanical engineering, the research offered a deeper look at how heat exchangers perform in the real world, where complex factors shape outcomes. “There are lots of textbook examples about heat exchangers,” he notes. “But to build the most efficient system, what matters is not how heat exchangers behave in theory — it’s how they perform in your house or your car.”

Download the paper: “Machine learning based prediction of airflow maldistribution in air-to-refrigerant heat exchangers.”

Published August 19, 2026