A surrogate modeling framework for learning spatiotemporal physical fields, powered by deep neural networks (e.g., Recurrent U-Net), and designed for CO₂ plume prediction.
- ✅ Supports 4D structured inputs using hierarchical deep learning model:
(B, T, C, X, Y, Z) - ✅ Modular training framework (Trainer class)
- ✅ Loss support (SSIM, Gradient, Perceptual)
simple_runet/
├── __init__.py
├── losses.py # MultiFieldLoss family
├── trainer.py # Trainer class
├── unet.py # U-Net & RUNet definitions
├── registry.py # Loss registration
├── utils.py # Misc tools
├── lpips.py
├── pretrained_networks.py
├── get_kernels_3d.py
├── requirements.txt
└── README.md
Run Case1-Train.ipynb for a simple 3D toy example.
(To Do ...) Run Case2-Train.ipynb for a simple 2D toy example.
torchtorchvisionkorniamatplotlibnumpy