BoLiu-USTC / HGPTrans

# HGPTrans: Hierarchical Graph-Pooling Transolver for Vehicle Drag Prediction
> Official PyTorch implementation of HGPTrans: hierarchical graph-pooling Transolver for vehicle aerodynamic drag prediction (Cd regression) from STL meshes.
Predict the aerodynamic drag coefficient (Cd) of a vehicle directly from its
surface mesh. The model builds a graph from the triangle mesh (vertices +
normals as node features, mesh edges as the graph edge set), runs GIN message
passing, applies Transolver-style physics attention over learned slice tokens
(suited to irregular meshes), and performs hierarchical TopK graph pooling
with a multi-scale pooled readout.
## Repository structure
 [code] 
## Installation
Requires Python 3.9+ and CUDA (for training on GPU).
 [code] 
`torch` and `torch_geometric` are listed but their CUDA-matched wheels depend
on your local CUDA version, so install them first following the official
guides if the pinned wheels do not match your driver:
- PyTorch: https://pytorch.org/get-started/locally/
- PyG: https://pytorch-geometric.readthedocs.io/en/latest/install/installation.html
## Data layout
Point `--data_dir` at a dataset directory with this layout
(DrivAerNet / DrivAerNet++ style):
 [code] 
The DrivAerNet++ 8k+ variant (`train.txt` / `val.txt` / `test.txt` splits and
`DrivAerNetPlusPlus_Cd_8k_Updated.xlsx`) is auto-detected as well.
## Training
 [code] 
Key options: `--accumulation_steps`
(default 16, i.e. effective batch = batch_size × accumulation_steps =
16 × 1). Vertex features are standardized per sample (zero mean / unit
variance over each mesh).
## Evaluation
 [code] 
Per-design predictions are written to `predLog/preds.txt` next to the
checkpoint; metrics (MSE / MAE / R² / rel-L2 / Pearson / Spearman) are printed
and an error-scatter plot is saved to `predLog/error.png`.
## Notes
- Keep `--batch_size 1`: model forwards use dense reshapes that assume a
 single graph per batch. Use `--accumulation_steps` for larger effective
 batches.
- All meshes above `--simplify_size` vertices (default 417,114) are decimated
 with quadric edge collapse before graph construction.
- Training latency printed during validation is forward-only; the end-to-end
 per-car inference time (mesh loading + QEM decimation + forward) reported in
 the paper is measured over the full pipeline.
## Citation
The accompanying paper is available on arXiv:2609.31765.
If you use this code or method, please cite:
 [code] 
## License
Released under the MIT License.