AI predicts water droplet spreading dynamics and icing patterns on cold surfaces

From tiny raindrops that patter against windowpanes to larger drops that freeze onto power lines or aircraft wings, the impact of aqueous droplets on cold surfaces is a complex physical process. When a droplet makes contact with a cold surface, it spreads maximally against the surface while heat transfer facilitates its rapid temperature reduction and potential freezing.

Understanding such dynamics can reduce potential hazards and optimize various industrial processes. But so far, it has been a challenge to accurately predict the spreading dynamics and icing pattern of a water droplet, hindering advancement of icing-control and anti- and de-icing technologies.

Yang et al. developed two novel AI models to respectively predict the spreading dynamics and to classify icing patterns and corresponding surface supercooling degrees.

“Artificial intelligence and machine learning algorithms, as an emerging strategy used increasingly in fluid-based problems, can effectively improve prediction accuracy of droplet impact dynamics and icing patterns, as our work demonstrates,” said co-author Xin Zhong.

For droplet spreading, the authors employed the Back Propagation Neural Network method, a multi-layer feedforward network algorithm, to achieve a high prediction accuracy of 95.96%.

For icing profiles, the researchers used clustering vector quantization methods, and the deep learning algorithms of Convolutional Neural Networks to predict the surface supercooling degrees for specific icing profiles with accuracy as high as 90.57%.

“We hope this work can provide new ideas and insights for further studies on droplet dynamics and anti/de-icing, and offer guidance for aircraft anti-icing, spray-freeze drying, and influenza vaccines storage,” said Zhong.

Source: “BPNN and CNN-based AI modelling of spreading and icing pattern of a water droplet impact on a supercooled surface,” by Song Yang, Yu Hou, Yuheng Shang, and Xin Zhong, AIP Advances (2022). The article can be accessed at http://doi/org/10.1063/5.0082568 .