Recurrent graph convolutional multi-mesh autoencoder for unsteady transonic aerodynamics

Unsteady, high-fidelity aerodynamic load predictions around a three-dimensional configuration will remain computationally expensive for the foreseeable future. Data-driven algorithms based on deep-learning are an attractive option for reduced order modelling of complex, nonlinear systems. However, a dedicated approach is needed for applicability to large and unstructured domains that are typical in engineering. This work presents a geometric-deep-learning multi-mesh autoencoder framework to predict the spatial and temporal evolution of aerodynamic loads for a finite-span wing undergoing different types of motion. The novel framework leverages on: (a) graph neural networks for aerodynamic surface grids embedded with a multi-resolution algorithm for dimensionality reduction; and (b) a recurrent scheme for time-marching the aerodynamic loads. The test case is for the BSCW wing in transonic flow undergoing a combination of forced-motions in pitch and plunge. A comprehensive comparison between a quasi-steady and a recurrent approach is provided. The model training requires four unsteady, high-fidelity aerodynamic analyses which require each about two days of HPC computing time. For any common engineering task that involves more than four cases, a clear benefit in computing costs is achieved using the proposed framework as an alternative predictive tool: new cases are computed in seconds on a standard GPU.