Dataset for the paper "Predicting Transonic Flowfields in Non–Homogeneous Unstructured Grids Using Autoencoder Graph Convolutional Networks"

This repository provides the Computational Fluid Dynamics (CFD) reference datasets used in:G. Immordino, A. Vaiuso, A. Da Ronch, M. Righi."Predicting transonic flowfields in non–homogeneous unstructured grids using autoencoder graph convolutional networks."Journal of Computational Physics, Vol. 524, 113708, 2025.https://doi.org/10.1016/j.jcp.2024.113708 The datasets contain steady Reynolds-averaged Navier–Stokes (RANS) simulations of transonic aerodynamic flows on unstructured grids, used for training and testing a Gradient-Based Autoencoder Graph Convolutional Network surrogate model. Two benchmark configurations of increasing geometric and physical complexity are included: Benchmark Supercritical Wing (BSCW) Semi-span wing with rectangular planform and supercritical airfoil. Grid: unstructured, ~86.8k surface points. Outputs: surface pressure coefficient (CP), skin friction coefficient components (CFx,CFy,CFz). NASA Common Research Model (CRM) Full wing–body geometry representative of a wide-body transport aircraft. Grid: unstructured, ~78.8k surface points. Outputs: surface pressure coefficient (CP), skin friction coefficient components (CFx,CFy,CFz). Generation Details Solver: SU2 v7.2.1, RANS formulation with Spalart–Allmaras turbulence model. Convergence criterion: Cauchy residual method on CL (tolerance 10−7). Discretisation: JST central scheme with artificial dissipation; gradients via Green–Gauss method. Sampling: 70 design points generated via Latin Hypercube Sampling (LHS). Flight envelope: Mach number = [0.70 , 0.84] , angle of attack AoA = [0 , 5] deg. PurposeThese datasets were used to train, validate, and benchmark autoencoder graph convolutional networks for predicting steady transonic flowfields on non-homogeneous unstructured grids. Applications include aerodynamic load estimation, surrogate modelling, and reduced-order modelling for uncertainty quantification. Contents CFD snapshots of surface pressure and skin-friction fields at each Mach–AoA condition for the BSCW and CRM test cases. Mesh connectivity and node coordinate files for graph representation. KeywordsTransonic aerodynamics; unstructured grids; geometric deep learning; graph neural networks; autoencoder; CFD; Benchmark Supercritical Wing; NASA CRM; reduced-order modelling; flowfield prediction.