Skip to main content
Back to timeline
arXivSource publication:

Airfoil surrogates drift off the optimization path: multi-step residual correction cuts CL error 59% for 0.25 extra seconds per sample

Synopsis

Aerodynamic surrogate models are efficient inside the training distribution but lose accuracy sharply on out-of-distribution geometries. A residual-guided sequential inference framework first produces an initial near-wall flow field with a Flow Initializer, then alternates correctors aimed at high-residual regions and shock regions, rebuilding only the masked local parts. On an out-of-distribution test set built from a CFD aerodynamic optimization trajectory, L1, CL and CD errors fall by 30%, 59% and 41% relative to single-step Geo-to-Flow prediction, at about 0.25 seconds of extra inference per sample.

Source-provided article image: Residual-Guided Multi-Step Conditional Generation for Fast Airfoil Flow Field Prediction
arXiv

Interpretation

Near-wall airfoil flow prediction on out-of-distribution geometries no longer rests on a single end-to-end mapping; the initial prediction becomes a starting point that can be locally rewritten and rolled back. The usual single-step Geo-to-Flow prediction is accurate within the training distribution, but for airfoils along an optimization trajectory whose geometry is not explicitly constrained, the initial flow field is often unreliable and structures such as shocks deviate clearly from ground truth. On the out-of-distribution test set built from a RAE2822 aerodynamic optimization trajectory, the single-step initial field has L1 error 6.71e-03, CL MAE 6.87e-02 and CD MAE 2.85e-03; after multi-step iteration these become 4.69e-03, 2.80e-02 and 1.68e-03, reductions of 30%, 59% and 41%.

The correction rewrites only the untrustworthy parts of the flow field and keeps the rest as visible conditioning. All three components share one near-wall Conditional Transformer backbone and differ only in how training inputs, outputs and masks are constructed, not in network architecture. At inference the near-wall residual field is evaluated first, and the top 50% of cells by residual are masked and reconstructed by the Residual-based Corrector; then the top 8% pressure-jump region, expanded by 8 cells on each side, is reconstructed by the Shock-based Corrector using the non-shock region as visible condition.

Iteration is governed by a set of stopping rules: it rolls back and terminates when the residual rises, treats an unchanged flow field as a fixed point and stops, and imposes a maximum iteration count to prevent infinite recursion. After each full alternating round a stopping criterion is checked; if the residual rises, the update is rejected and the solution rolls back to the lowest-residual result of the previous step, rather than running a fixed number of steps. On the out-of-distribution test set the average iteration count is 3.47 with a median of 3, and per-sample inference rises from 0.061 s for the single step to 0.309 s, about 0.25 s extra per sample, measured on an RTX4060 at batch size 1.

Perspective

The results cover two-dimensional airfoil near-wall flow fields on a grid of 296 chordwise cells and two wall-normal cell layers, with the test set made of candidate airfoils along a CFD-solved aerodynamic optimization trajectory starting from RAE2822. Training used 45,000 supercritical airfoil samples with CFL3D high-fidelity flow fields, interpolated onto uniform algebraic grids on which the surrogates run directly. The authors propose extending the framework from two-dimensional airfoil flows to three-dimensional aerodynamic problems.

A few extremely difficult out-of-distribution samples still retain some shock-structure error after iteration, though the flow field is clearly more accurate than the initial prediction. The correctors' validation errors come from an in-distribution validation set and do not fully match their inference-time input setting. The authors expect further cost reductions once the iterative framework runs in a multi-batch parallel setting.