{
  "domain": "engineering-energy",
  "image": "python:3.12-slim",
  "data": "The PDEBench and The Well training splits; the test trajectories are scored on the runner.",
  "benchmark": {
    "task": "PDE surrogate modelling",
    "metric": {
      "name": "nRMSE",
      "better": "lower"
    },
    "cost": "GPU-hours",
    "subtasks": "systems",
    "guardrails": [
      "Autoregressive rollout from the first ten steps, with no ground truth after them",
      "One model per system, trained on its training split only",
      "At most 48 GPU-hours per system"
    ],
    "aggregate": "mean",
    "baseline": {
      "name": "FNO",
      "score": 0.091,
      "cost": 9.5,
      "subtasks": {
        "1D advection": 0.1281,
        "1D Burgers": 0.1477,
        "1D diffusion–reaction": 0.09651,
        "2D Darcy flow": 0.05843,
        "2D shallow water": 0.1251,
        "2D reaction–diffusion": 0.01769,
        "2D compressible NS": 0.1182,
        "3D compressible NS": 0.03628
      }
    }
  }
}