Aeroelastic coupling stops rerunning CFD every iteration: full-field physics models converge a transonic swept wing in 8 to 11 minutes on one GPU

The most expensive step of an aeroelastic coupling loop, a CFD solve repeated at every iteration, can be replaced by a full-field physics model that predicts pressure and wall shear stress over the whole wing surface directly from CAD geometry, cutting the loop from hours to minutes. Previously each coupling iteration crossed a team and tool boundary to request a flow solution, and engineers often skipped coupled analysis in early design, so the interaction surfaced late as rework. On the static aeroelastic case of a transonic swept wing, the CFD loop with mesh adaptation on three H100 GPUs took about four hours on average to converge; the SHIFT-Wing plus TACS loop converged in 8 to 11 minutes on one T4, which the source calls more than 20 times faster.

No meshing step remains in the loop, and the deformation returns to the model as new CAD, with loads and displacements transferred conservatively between wing surface and wing box. The usual route exports one solver's solution, maps it onto the other solver's mesh and applies it as a load, often remeshing the deformed shape before the next solve, and automating that chain is a project of its own. The structure is solved with the open-source finite element solver TACS on a shell model of the wing box, loads and displacements pass between wing surface and wing box through MELD, and skin and spar thicknesses vary along the span but are fixed across cases.

The structural solver can also be replaced by a second full-field physics model, so two models exchange full fields each iteration and no conventional solver remains in the loop. The first step keeps the structural solver, the setup for a structures engineer who wants to keep their own solver and take loads from an aerodynamic model the aerodynamics team has signed off on. SHIFT-Wingbox, a modified Luminary-SMART architecture that takes both geometry and the surface fields on it, was trained on about 1,400 TACS solutions; paired with SHIFT-Wing on three wings, its converged washout lift loss agreed within 1 percentage point and tip twist within 0.04 degrees of the loop containing TACS.

The demonstration is limited to the static aeroelastic response of a transonic swept wing, showing washout and the resulting lift loss on the SHIFT-Wing geometry and treating aeroelastic twist as the main geometric quantity that changes between coupling iterations. The source states that bending is ignored as a second-order effect, but the same analysis can be extended to include flight conditions and the associated bending deformations when they are nontrivial.

The source notes that most of the difference in absolute lift is already present at the undeformed shape, coming from SHIFT-Wing's prediction error on each geometry at about 1% to 2% of the undeformed lift. The demonstration covers static aeroelastic twist only, with bending treated as a second-order effect.