Update README.md

README.md

@@ -1,3 +1,134 @@
----
-license: mit
----
+---
+license: cc-by-4.0
+task_categories:
+  - other
+language:
+  - en
+tags:
+  - seismic
+  - structural-engineering
+  - ground-motion
+  - hdf5
+  - time-series
+pretty_name: KNET Seismic Ground Motion & Building Response Dataset (MDOF)
+size_categories:
+  - 10M<n<100M
+---
+
+# KNET Seismic Ground Motion & Building Response Dataset (MDOF)
+
+Ground motion records from the K-NET strong-motion network (Japan), paired with numerically simulated floor acceleration responses for 250 multi-degree-of-freedom (MDOF) shear-building configurations. Used to train and evaluate Fourier Neural Operator (FNO) models for seismic structural response prediction.
+
+## Dataset Summary
+
+| Property | Value |
+|----------|-------|
+| Ground motion records | 3,474 (K-NET) |
+| Amplitude scale factors per GM | 57 |
+| Building configurations | 250 MDOF shear buildings |
+| Signal length | 3,000 time steps (60 s @ 50 Hz) |
+| GM file format | HDF5 (`.h5`) |
+| Building response format | HDF5 (`.h5`), one file per building |
+
+## File Structure
+
+```
+MDOF/
+├── All_GMs/
+│   └── GMs_knet_3474_AF_57.h5        # Ground motion inputs
+└── knet-250/
+    └── Data/
+        └── fno/
+            ├── Blg_F6_18m_IM7_st0.h5.h5       # Floor acc. response
+            └── ...
+```
+
+### Ground Motion File (`GMs_knet_3474_AF_57.h5`)
+
+Each entry stores the scaled acceleration time series for one GM × scale-factor combination.
+
+| HDF5 Key pattern | Shape | Description |
+|-----------------|-------|-------------|
+| `gm_{i}/af_{j}/data` | `(3000,)` | Scaled acceleration (m/s²), 50 Hz |
+| `gm_{i}/af_{j}/pga` | scalar | Peak ground acceleration (m/s²) |
+| `gm_{i}/metadata` | attrs | Station, event, magnitude, etc. |
+
+### Building Response Files (`Blg_F6_18m_IM7_st0.h5.h5`)
+
+Each file stores the simulated structural response for all GM × scale combinations for one building.
+
+| HDF5 Key | Shape | Description |
+|----------|-------|-------------|
+| `response/gm_{i}/af_{j}/floor_acc` | `(n_floors, 3000)` | Floor acceleration (m/s²) |
+| `building/attributes` | attrs | Number of floors, periods, damping, etc. |
+| `building/damage_state/gm_{i}/af_{j}` | scalar int | HAZUS damage state (0–4) |
+
+## Loading
+
+### Option 1 — Direct HDF5 access (h5py)
+
+```python
+import h5py
+import numpy as np
+
+GM_FILE = "path/to/MDOF/All_GMs/GMs_knet_3474_AF_57.h5"
+
+with h5py.File(GM_FILE, "r") as f:
+    # Load ground motion i=0, amplitude factor j=0
+    gm = f["gm_0/af_0/data"][:]          # shape (3000,)
+    pga = f["gm_0/af_0/pga"][()]
+
+print(f"GM shape: {gm.shape}, PGA: {pga:.4f} m/s²")
+```
+
+```python
+import h5py
+import numpy as np
+
+BLG_FILE = "path/to/MDOF/knet-250/Data/fno/building_0001.h5"
+
+with h5py.File(BLG_FILE, "r") as f:
+    # Floor acceleration response for GM i=0, AF j=0
+    floor_acc = f["response/gm_0/af_0/floor_acc"][:]   # (n_floors, 3000)
+    damage_state = f["building/damage_state/gm_0/af_0"][()]
+
+print(f"Floor acc shape: {floor_acc.shape}, Damage state: {damage_state}")
+```
+
+### Option 2 — PyTorch Dataset (recommended for training)
+
+Clone the [SeismicFNO](https://github.com/HKUJasonJiang/Seismic-FNO) repository and use `DynamicDataset`:
+
+```python
+import numpy as np
+from torch.utils.data import DataLoader
+from module.dataprep_v2 import DynamicDataset
+
+GM_FILE      = "path/to/MDOF/All_GMs/GMs_knet_3474_AF_57.h5"
+BUILDING_DIR = "path/to/MDOF/knet-250/Data/fno/"
+
+# Full dataset (all 3474 × 57 combinations)
+dataset = DynamicDataset(
+    gm_file_path       = GM_FILE,
+    building_files_dir = BUILDING_DIR,
+)
+
+# Or pass a pre-computed index array for train/val/test splits
+rng     = np.random.default_rng(42)
+indices = rng.permutation(len(dataset))
+train_ds = DynamicDataset(GM_FILE, BUILDING_DIR, gm_indices=indices[:int(0.7 * len(indices))])
+
+loader = DataLoader(train_ds, batch_size=64, shuffle=True, num_workers=4)
+
+# Each batch: (gm, building_attributes, floor_acc_response, damage_state)
+gm, attr, resp, ds = next(iter(loader))
+print(gm.shape, resp.shape)   # (64, 3000, 1), (64, 3000, 1)
+```
+
+## Citation
+
+If you use this dataset, please cite the K-NET strong-motion network and the associated SeismicFNO paper (forthcoming).
+
+## License
+
+[Creative Commons Attribution 4.0 (CC BY 4.0)](https://creativecommons.org/licenses/by/4.0/)