On real-time management of on-board ice protection systems by means of machine learning

Machine learning tools are applied to hasten the numerical prediction of ice formation on local portions of aircraft flying in hazardous weather conditions, for a broad flight envelope. In aeronautical applications, the numerical prediction of ice accretion is a challenging multi-physics problem which is generally investigated by solving a sequence of computationally demanding steps. Conventionally, numerical simulations are carried out to reconstruct the relative trajectory of droplets suspended in the atmosphere and to ultimately predict the water impinging points over the aircraft. This procedure requires the evaluation of an expensive mathematical model and it involves a Lagrangian particle tracking software coupled to a computational fluid dynamics solver. Here, different machine learning classification algorithms are trained on the synthetic data set generated using the full computational framework. The diverse algorithms are eventually compared in terms of efficiency and accuracy. Note that, once available, the synthetic data set is expected to be substituted by a comprehensive data set of real experimental measurements. In-flight ice protection requires the activation of on-board systems which absorb a significant amount of power. A cheap and reliable ice prediction capability could be of particular interest to optimize the real-time management of modular anti-ice systems, such as electro-thermal devices, potentially leading to considerable energy savings during a typical mission. This work embodies a proof of concept for the future improvement of real-time management for ice protection systems. The potential of this approach is not limited to aerospace applications but it may be relevant for wind-energy and naval applications.