Pretrained aero surrogates degrade on a new car class: GeoReplay cuts source error in 77 of 80 settings
Synopsis
Neural surrogates for external vehicle aerodynamics are usually trained and evaluated on a single vehicle dataset, and fine-tuning one on a new vehicle class often loses to training from scratch while damaging its accuracy on the original data. GeoReplay replays source simulations weighted by geometric proximity during adaptation and keeps a source gradient only when it agrees with the target gradient. Across five DrivAerML-pretrained surrogates, four independent target datasets and several budgets, 80 settings in all, it lowers the fine-tuned source pressure error in 77 while matching target accuracy.
Interpretation
When a pretrained aerodynamic surrogate moves to a new vehicle class, ordinary fine-tuning does not reliably beat training from scratch and sharply raises its error on the original vehicle dataset. Most automotive surrogates had been trained and evaluated within one vehicle dataset, and transfer across independent CFD datasets had not been measured systematically; after fine-tuning, source error rose far above the zero-shot level, for example from 0.042 to 0.350 for Transolver adapted to DrivAerNet++. In 40 pressure and wall-shear comparisons, fine-tuning beat training from scratch in only 17 and was worse in 23; across 80 combinations of surrogate, target dataset and budget, GeoReplay lowered the fine-tuned DrivAerML pressure error in 77 settings, with a median target error close to that of fine-tuning.
Keeping source simulations as replay data during adaptation can hold both target and source accuracy. Plain experience replay reweights source simulations uniformly, geometry weighting alone selects source simulations by proximity, and gradient alignment alone keeps only source updates that agree with the target gradient; either component alone raises target error on some targets. In the Transolver and GeoTransolver component ablation, plain replay with uniform weights cut DrivAerML error by roughly 3.6 to 4.9 times, but both the plain update and the aligned update with uniform weights raised target error on SHIFT-SUV; only with geometry weighting and gradient alignment combined did target error stay within 10% of fine-tuning on every target dataset.
Perspective
The conclusions rest on five surface surrogates pretrained on DrivAerML and four target datasets, DrivAerNet++, SHIFT-SUV, AhmedML and WindsorML, with errors as relative pressure and wall-shear errors and the main comparison at ten target training simulations.
Variance over three seeds was measured only at ten target training simulations on one split per target, so small differences at other budgets need caution; the geometric measures omit CFD solver, turbulence model, operating-condition, mesh and label differences that may affect transfer.
