Model Checkpoints for "Cross-Domain Pretraining for Steady-State CFD Surrogates"
Pretrained checkpoints for models used in the paper. Each folder (AB-UPT/SMART/Transolver) contains models implemented in the Github and trained on different datasets
In general, the released checkpoints do quite well (achieve relatively low errors and capture relevant features on the validation sets). For downstream applications (i.e., deploying to a new dataset), it may be of interest to use a pretrained, joint model, although this fixes the downstream architecture to AB-UPT, Transolver, or SMART. However, the benefit of this is that you may need to generate less data for the new application, or reach a lower error compared to training from scratch, if the dataset is already generated.
License
The models are trained on Emmi Wing and DrivAerNet++, which have a non-commercial license. Therefore using any of the pretrained models for commercial uses is prohibited.