{ "model_name": "SEEDS", "model_type": "seeds", "architectures": [ "SEEDS", "SEEDSModel" ], "framework": "PyTorch", "domain": "atmosphere", "task": "conditional-diffusion-ensemble-weather-forecasting", "implementation": { "entry_point": "model/seeds.py", "scope": "Conditional score network for cubed-sphere weather fields; the local data contract is compatible with the project scripts and does not replace official GEFS/ERA5 preprocessing" }, "architecture": { "family": "conditional diffusion score network with axial attention over spatial patches, variables, and forecast snapshots", "input_format": "BCFHW", "seed_input_format": "BSCFHW", "output_format": "BCFHW", "channels": 8, "faces": 6, "face_grid_shape": [ 48, 48 ], "seed_count": 2, "target_ensemble_count": 16, "patch_size": 12, "embed_dim": 768, "spatial_layers": 6, "field_layers": 4, "sequence_layers": 6, "mlp_ratio": 4, "dropout": 0.0, "prediction": "score_noise", "noise_schedule": { "sigma_min": 0.01, "sigma_max": 100.0, "sigma_formula": "sigma(t) = sigma_min * (sigma_max / sigma_min) ** t" }, "sampling": { "default_steps": 16, "configured_solver": "euler_maruyama", "local_implementation": "explicit reverse sigma-schedule update without additional stochastic noise between steps" } }, "data": { "dataset": "GEFS-conditioned weather ensemble data with ERA5 evaluation reference", "variables": [ "mean_sea_level_pressure", "temperature_2m", "eastward_wind_850hpa", "northward_wind_850hpa", "geopotential_500hpa", "temperature_850hpa", "total_column_water_vapour", "specific_humidity_500hpa" ], "lead_time_days": 7, "normalization_note": "Scientific evaluation requires the matching anomaly/climatology normalization and denormalization; local synthetic data only checks tensor and pipeline contracts." }, "configuration_sources": [ "conf/config.yaml", "model/seeds.py", "scripts/common.py", "scripts/data_loader.py" ] }