{
  "model_name": "GraphDOP",
  "model_type": "graphdop",
  "architectures": [
    "GraphDOP"
  ],
  "framework": "PyTorch",
  "domain": "climate-and-atmosphere",
  "task": "observation-driven-medium-range-weather-forecasting",
  "implementation": {
    "entry_point": "model/graphdop.py",
    "scope": "pure-PyTorch minimal reproduction using gridded ERA5 placeholders and fixed regular-mesh graphs instead of the paper's irregular Level-1 observations and dynamic graphs"
  },
  "architecture": {
    "family": "GNN encoder-Transformer processor-GNN decoder",
    "input_format": "B T C H W",
    "encoder": "per-grid-cell MLP, adaptive pooling to the latent mesh, then residual mean-aggregation GNN layers",
    "processor": "pre-normalized Transformer encoder over latent-mesh tokens with learned positional embeddings",
    "decoder": "latent-mesh GNN, bilinear upsampling, and a per-grid-cell output MLP",
    "edge_features": [
      "forward bearing",
      "Haversine distance"
    ],
    "activation": "GELU",
    "normalization": "LayerNorm",
    "loss": "channel-weighted mean squared error",
    "repository_default_config": {
      "purpose": "connectivity validation with synthetic gridded data",
      "grid_shape": [
        32,
        32
      ],
      "mesh_shape": [
        8,
        8
      ],
      "in_channels": 6,
      "out_channels": 6,
      "input_steps": 2,
      "output_steps": 2,
      "latent_dim": 64,
      "num_encoder_layers": 2,
      "num_decoder_layers": 2,
      "num_processor_blocks": 1,
      "attention_heads": 4,
      "hidden_dim": 64,
      "channel_weights": [
        1,
        1,
        1,
        1,
        1,
        1
      ]
    },
    "paper_reference_config": {
      "latent_grid": "O96 reduced Gaussian grid with 40320 nodes",
      "latent_dim": 1024,
      "observation_graph": "dynamic graph over irregular Level-1 observations",
      "training_steps": 70000,
      "training_hardware": "64 H100 GPUs"
    }
  },
  "data": {
    "dataset": "ERA5",
    "role": "regular-grid placeholder for the paper's multi-instrument observations",
    "variables": [
      "atms_brightness_temperature",
      "gpsro_bending_angle",
      "ascat_sigma0",
      "significant_wave_height",
      "2m_temperature",
      "10m_wind_speed"
    ],
    "time_step_hours": 6,
    "input_length": 2,
    "output_length": 2,
    "channels": 6,
    "spatial_size": [
      32,
      32
    ],
    "storage_format": "HDF5 fields with T C H W layout",
    "train_years": [
      1951,
      1952
    ],
    "validation_years": [
      1953
    ],
    "test_years": [
      1954
    ]
  },
  "configuration_sources": [
    "README.md",
    "conf/config.yaml",
    "model/graphdop.py",
    "scripts/train.py",
    "scripts/inference.py",
    "scripts/fake_data.py",
    "configuration.json"
  ]
}