The study, published in Frontiers in Thermal Engineering, tackled a phenomenon known as condensation frosting, a multi-stage process in which water vapor from moist air first nucleates into tiny droplets, then grows, merges, and finally freezes into a solid sheet of ice. Engineers have long known that surface texture influences every one of these stages by altering wettability and droplet pinning, the tendency of a droplet’s contact line to anchor itself to surface defects. But pinning down the precise relationship between texture, condensation dynamics, and freezing speed has remained elusive, largely because a single microscope image of condensing droplets can contain hundreds or even thousands of individual drops that no human could reliably count by hand.

That counting problem is where the AI comes in. The team, led by Jane C. Turner and Amy R. Betz of the Alan Levin Department of Mechanical and Nuclear Engineering, together with computer scientists Aryan Singh Dalal and Hande Küçük McGinty, and Melanie M. Derby, built an object detection model based on the YOLOv11 architecture, a deep convolutional neural network designed to detect and classify objects in real time. The model was trained on more than 1,300 manually annotated microscope images, with undergraduate students spending countless hours drawing bounding boxes around individual droplets. After 110 training epochs, using the Adam optimizer with an initial learning rate of 0.001 and a batch size of 16, the network achieved a mean average precision of 78 percent at an intersection-over-union threshold of 0.50, with precision of 75 percent and recall of 76.5 percent. In practical terms, the model could identify droplets as small as 4.5 micrometers and track their number, average diameter, phase state, and surface coverage across entire fields of view.

The experimental setup was elegantly simple. A bar of 6061 aircraft-grade aluminum with a mirror finish served as the smooth control surface, while a second sample was patterned with a femtosecond laser operating at a wavelength between 800 and 1,056 nanometers, a pulse duration of 184 femtoseconds, and a repetition rate of 50 kilohertz. The laser carved parallel lines 100 micrometers wide, spaced 150 micrometers apart, and redeposited material into ridgelines, or peaks, that rose an average of 6.8 micrometers above the surface, with individual peaks ranging from 3.3 to 9.5 micrometers tall. Goniometer measurements revealed the control surface was hydrophobic with a contact angle of 94 degrees, while the microtextured surface displayed striking directional wetting: 129 degrees when the line pattern aligned with the lens, but only 96.7 degrees when the lines ran perpendicular to it.

Experiments took place in an environmental chamber at Kansas State University’s Institute for Environmental Research, with ambient air temperatures between 20 and 23 degrees Celsius and relative humidity between 39 and 47 percent. Each aluminum sample was mounted on a Peltier cooler set to minus 10 degrees Celsius, driving the surface to minus 8 degrees Celsius within 20 to 30 seconds, cold enough for condensation from the moist air and subsequent freezing. A Leica DVM2500 microscope recorded video at 600 times magnification, yielding a field of view of just 0.51 by 0.38 millimeters, small enough to resolve the initial nucleation of individual droplets. Embedded K-type thermocouples, calibrated to an accuracy of plus or minus 0.5 degrees Celsius, logged surface temperature at 1 hertz throughout each run.

When the researchers fed the resulting videos, reduced to one frame per second, through the neural network, a clear pattern emerged. On both surfaces, the droplet population and average droplet diameter followed power-law scaling over time, meaning the population declined in proportion to time raised to a negative exponent while diameter grew in proportion to time raised to a positive exponent. But the exponents differed measurably between the two surfaces. On the smooth control, the population declined with an exponent of minus 1.5 and diameter grew with an exponent of 0.63, consistent with coalescence-dominated growth, in which droplets merge and shed, re-nucleating new droplets in the voids left behind. On the microtextured surface, the population exponent was only minus 1.3 and the diameter exponent 0.57, closer to the value of 0.5 expected for diffusion-dominant growth, in which droplets grow primarily by direct vapor deposition rather than merging.

That shift in exponents is the study’s central technical finding. The ridgelines on the laser-patterned surface increased droplet pinning and solid-liquid contact area, immobilizing droplets and suppressing coalescence. With fewer droplets merging, the population declined more slowly and the growth regime transitioned earlier toward direct condensation. The researchers interpret this as surface texture fundamentally altering the balance between vapor deposition and coalescence, shifting condensation dynamics in a way that could be captured by nothing more than two numbers extracted from the AI model’s output.

And those two numbers predicted freezing with remarkable fidelity. The microtextured surfaces achieved full frost coverage in an average of 139.5 seconds, roughly three times faster than the smooth control surfaces, which took an average of 419.5 seconds. Individual trials showed the gap ranging from 1.8 times to 3.9 times. The team attributes the accelerated freezing not to any intrinsic change in ice growth rate but to the altered condensation dynamics: increased pinning and contact area lowered the energy barrier for heterogeneous nucleation and improved heat transfer from droplet to substrate. Intriguingly, ice bridging, in which ice dendrites propagate from frozen droplets to adjacent supercooled liquid droplets, dominated the freezing front propagation on both surfaces, meaning the texture changed how fast frost arrived, not how it spread.

The AI model was not flawless. During the first seconds of nucleation, when frames can contain over 1,800 tiny droplets, the network struggled to identify every drop, partly because training images of initial nucleation were scarce and time-consuming to annotate. On the microtextured surface, the model occasionally mistook regions of surface texture for droplets, an error that diminished after the initial nucleation phase. Confidence hovered around 70 percent for both surfaces during the well-annotated middle portion of each experiment, and the researchers expect performance to improve across the entire timeline as the dataset grows.

The implications reach far beyond the laboratory bench. By establishing that surface-induced pinning directly governs both condensation growth and subsequent freezing behavior, the study gives engineers a quantitative design tool: if you want to delay icing on heat exchangers, aircraft wings, or power infrastructure, the power-law exponents extracted by an AI model tell you whether your surface is promoting the coalescence-driven growth that keeps droplets mobile, or the pinned, direct-growth regime that accelerates frost. The team notes that the YOLO workflow can be adapted to related multiphase phenomena, from spray-pattern characterization to bubble nucleation during boiling, and that future work integrating neuro-symbolic AI approaches could improve the explainability of results across a broader range of engineered surfaces and substrate materials. In the contest between ice and infrastructure, it turns out the most powerful weapon may be a laser, a microscope, and a neural network that never gets tired of counting droplets.

Subject of Research: AI neural network analysis of droplet condensation and freezing on laser-patterned aluminum reveals that surface pinning shifts power-law growth exponents and triples freezing speed compared to smooth surfaces.

Article Title: Examining condensation and freezing behavior on a laser-patterned metallic surface using an AI neural network model

Article References: Turner, J. C., Dalal, A. S., McGinty, H. K., Derby, M. M., & Betz, A. R. (2026). Examining condensation and freezing behavior on a laser-patterned metallic surface using an AI neural network model. Frontiers in Thermal Engineering, 6, Article 1808591. https://doi.org/10.3389/fther.2026.1808591

Image Credits: AI Generated

DOI: 10.3389/fther.2026.1808591

Keywords: condensation frosting, droplet pinning, dropwise condensation, YOLOv11, neural network, laser surface texturing, aluminum, ice bridging, power-law scaling, machine learning, thermal engineering, freezing mechanisms