U-Net (3D U-Net)¶

RealPDEBench uses a 3D U-Net baseline (encoder–decoder CNN with skip connections) to model spatiotemporal fields.

Reference Paper: Ronneberger et al., MICCAI 2015.

@inproceedings{ronneberger2015u,
  title={U-net: Convolutional networks for biomedical image segmentation},
  author={Ronneberger, Olaf and Fischer, Philipp and Brox, Thomas},
  booktitle={International Conference on Medical image computing and computer-assisted intervention},
  pages={234--241},
  year={2015},
  organization={Springer}
}

RealPDEBench interface¶

• Input: x with shape [B, T_in, H, W, C_in]
• Output: predictions with shape [B, T_out, H, W, C_out]

Implementation notes (RealPDEBench)¶

• The implementation permutes inputs to [B, C_in, T_in, H, W] for 3D convolutions and permutes back before returning outputs.
• If \(T_\text{out} > T_\text{in}\), the baseline repeats the input along the time axis to reach T_out, so it assumes \(T_\text{out}\) is a multiple of \(T_\text{in}\).
• Training uses an MSE-style objective on the predicted fields.

U-Net specific YAML config¶

This baseline is enabled by model_name: "unet" in the training YAML.

Config files¶

• Cylinder: realpdebench/configs/cylinder/unet.yaml
• Controlled Cylinder: realpdebench/configs/controlled_cylinder/unet.yaml
• FSI: realpdebench/configs/fsi/unet.yaml
• Foil: realpdebench/configs/foil/unet.yaml
• Combustion: realpdebench/configs/combustion/unet.yaml

Model-specific keys¶

These keys are consumed by realpdebench.model.load_model.load_model() and realpdebench.model.unet.Unet3d.

• dim_mults (list[int]): Channel multipliers per U-Net resolution stage (controls width growth across downsampling/upsampling).