A Graph Neural Network Simulator for Transonic Vortex-Dominated Flows on the DLR-F23 Configuration

Kurzfassung

The high computational expense of Reynolds-Averaged Navier-Stokes (RANS) simulations creates a bottleneck in analyzing complex, vortex-dominated transonic flows. This study addresses this challenge by evaluating a Graph Neural Network (GNN) surrogate model to predict surface pressure distributions. A high-fidelity database of 180 RANS simulations of the DLR-F23 triple-delta-wing was generated using the CODA solver, clustered around Mach 0.85 and angles of attack of 12°, 15° and 18°. A GNN was trained on two-thirds of this data. Results demonstrate good predictive accuracy for interpolative cases within the training convex hull. However, performance degrades for extrapolation cases involving flow topology changes not seen in training. These findings suggest GNNs are highly promising for accelerating analysis within well-defined envelopes.