Snow, ice, and algorithms: AI serving aviation safety

An article by Valérie Levée, science journalist

In winter, if cold weather and snow do not ground airplanes, it is because they are sprayed with a mixture of water and propylene glycol or ethylene glycol to clear them of frost. The recipe for the mixture depends on weather conditions, and LIMA (International Anti-Icing Materials Laboratory / Laboratoire international des matériaux antigivres) is one of the few platforms in the world accredited to qualify de-icing and anti-icing fluids. Gelareh Momen holds the Institutional Research Chair on Innovative Anti-Icing Materials (Chaire institutionnelle de recherche en matériaux antigivre innovants - CiMAGI), is the former director of LIMA (2020-2025), and is a professor at the École de technologie supérieure. She has just published a study in the journal Aerospace Science and Technology showing how machine learning can help predict the effectiveness of these fluids.

In preparation for takeoff, a first orange fluid is sprayed onto the aircraft. This is the de-icing fluid, designed to remove snow, ice, and frost. While awaiting takeoff, a second fluid, often green, is applied to create a protective film that prevents new frost from forming. The effectiveness of this anti-icing protection decreases over time, a period known as holdover time. The pilot therefore has a restricted time window to take off. On the other hand, due to its viscous nature, anti-icing fluid can also potentially disrupt aerodynamics.

"The fluid must flow off effectively during the takeoff phase so as not to alter aerodynamic performance," mentions Gelareh Momen.

This is why fluids must adhere to strict standards both for holdover time and for fluid elimination from the aircraft's surface. However, holdover time depends on the type of fluid used, the aircraft model, and weather conditions such as temperature, as well as the nature and intensity of precipitation. The fluids must therefore be tested to determine the situations in which they are effective. Yet, these tests are costly, and reproducing the multitude of possible weather conditions is highly difficult. This explains the recourse to artificial intelligence.

AI enters the development of anti-icing materials

Gelareh Momen makes it clear that her expertise lies in chemical engineering and materials engineering, not artificial intelligence. However, she rode the AI wave and began integrating it into her research with a first study published in 2021 in the Chemical Engineering Journal. At the time, the goal was to predict the behavior of water droplets impinging on a superhydrophobic surface. Depending on the nature of the surface, will these droplets bounce, spread, or simply slide? AI makes it possible to model these types of complex non-linear phenomena to improve predictions. Since then, Gelareh Momen has continued to explore the potential of AI for her research projects, and applying AI to predict the holdover time of anti-icing fluids was a natural continuation.

Data to train AI

To train an algorithm, it must be supplied with data, and Gelareh Momen's team set to work testing anti-icing fluids in accordance with the Society of Automotive Engineers (SAE) ARP5485B standard. This involves an aluminum plate inclined at a 10-degree angle to simulate an aircraft wing. Such a setup was used to test the holdover time of fluids indoors under controlled conditions using LIMA's snow machine, and outdoors under natural snowfall conditions. The research team used this data to train three machine learning algorithms and compare their ability to predict fluid holdover time based on temperature and snow conditions. The results showed that ensemble models, particularly XGBoost, effectively capture the complex non-linear relationships between environmental variables and anti-icing fluid performance.

This is already an excellent performance, but to improve it, a larger volume of training data would be required to better reflect the high variability of weather conditions, especially extreme conditions. Testing under real or artificial conditions can hardly cover the entire spectrum of weather conditions, but here again, AI can come to the rescue with another type of algorithm using a generative model to produce synthetic data that can be added to the training data. This is the next step.

"The ultimate goal is not to be limited to experimental data, which is too costly, and to have an AI model that can predict the holdover time of a fluid regardless of temperature and precipitation rates," summarizes Gelareh Momen.

The researcher's recommendations

• In a national defense and security context, research investments are needed to ensure that infrastructure remains operational in northern climates. This is true for airplanes, but we must also consider wind turbines, helicopters, and drones.
• Currently, active methods, which rely on an energy supply, are widely used to control equipment icing. However, hybrid approaches, synergistically combining passive and active strategies, stand out as a promising solution. Passive methods, based on the development of icephobic materials capable of limiting snow and ice accumulation, help reduce energy requirements, while active systems ensure targeted control when necessary. Thus, this combination optimizes the overall performance of systems while significantly decreasing energy consumption and the carbon footprint of infrastructure operating in northern conditions. Artificial intelligence is an excellent tool for optimization, but validating the results remains essential.

Affiliations

Institut nordique du Québec

Québec Centre for Advanced Materials (Centre québécois des matériaux fonctionnels)

Research Center for High-Performance Polymer Systems and Composites (Centre de recherche sur les systèmes polymères et composites à haute performance)

To go further / Additional reading

Cardona, A., Nakouri, H., Jean-Denis Brassard, J.-D., & Momen, G. (2026).
Machine learning prediction of aircraft anti-icing fluid endurance: A comparative study using natural and artificial snow data. Aerospace Science and Technology, 176, Part A, 112103. https://doi.org/10.1016/j.ast.2026.112103

Keshavarzi, S., Entezari, A., Maghsoudi, K., Momen, G., & Jafari, R. (2022). Ice nucleation on silicone rubber surfaces differing in roughness parameters and wettability: Experimental investigation and machine learning–based predictions.
Cold Regions Science and Technology, 203, 103659. https://doi.org/10.1016/j.coldregions.2022.103659

Azimi Yancheshme, A., Enayati, S., Kashcooli, Y., Jafari, R., Ezzaidi, H., & Momen, G. (2022). Dynamic behavior of impinging drops on water repellent surfaces: Machine learning-assisted approach to predict maximum spreading. Experimental Thermal and Fluid Science , 139, 110743. https://doi.org/10.1016/j.expthermflusci.2022.110743

Keshavarzi, S., Sourati, J., Momen, G., & Jafari, R. (2022). Temperature-dependent droplet impact dynamics of a water droplet on hydrophobic and superhydrophobic surfaces: An experimental and predictive machine learning–based study,. International Journal of Heat and Mass Transfer, 195, 123190. https://doi.org/10.1016/j.ijheatmasstransfer.2022.123190

Azimi Yancheshme, A., Hassantabar, S., Maghsoudi, K., Keshavarzi,S., & Jafari, R. (2021) Integration of experimental analysis and machine learning to predict drop behavior on superhydrophobic surfaces. Chemical Engineering Journal, 417, 127898. https://doi.org/10.1016/j.cej.2020.127898