Geometric Deep Learning for Loads Prediction of Maneuvering Aircraft

A framework based on deep learning is developed for the rapid prediction of unsteady aerodynamic loads and flowfield distributions around a full-scale, maneuvering vehicle. The framework is demonstrated for the Future Fighter Demonstrator configuration, studied within the NATO Science & Technology Organisation Research Task Group 351, performing static and forced-motion maneuvers in pitch up to high angles of attack. The complexity of the problem is characterized by the dynamics of a set of vortices that form and grow in size and intensity until they break down, strongly dependent on the three-dimensional planform shape. For applicability to both structured and unstructured grids, we leverage geometric deep learning. For deployment on a large number of surface grid points, in the order of millions, we developed a multimesh scheme to operate at coarser levels of refinement. For robustness when predicting unsteady flow features, we proposed two schemes to limit the well-known pitfall of long-term error accumulation, namely a quasi-steady and a recurrent scheme. For efficiency in generating the model, training is limited to one single signal that provides sufficient coverage of the three-parameter space, namely static angle of attack, pitch motion amplitude, and frequency. Accuracy of the unsteady models is assessed on four forced-motion pitch signals and on predicting steady-state aerodynamic characteristics. We found the recurrent scheme to outperform the quasi-steady scheme across multiple error metrics. This is attributed to better capturing the flow hysteresis via inclusion of past flow information at each time step iteration.