# Эксперимент: Exact-Curl Neural Field Output (#930)

{
  "worked": false,
  "confidence": 9,
  "verdict": "Built a small periodic 2-D exact-curl CNN head and a direct-field CNN baseline with a divergence penalty. The numerical identity check gave max |D C A| = 2.22e-16, and the curl model's test divergence was only 1.19e-7 max / 1.76e-8 RMS versus 0.219 max / 0.0343 RMS for the baseline. However, the curl model had slightly higher field MSE (0.00377 vs 0.00350), so the experiment confirms exact constraint preservation but does not demonstrate an overall predictive win.",
  "metrics": {
    "baseline": "Direct 2-channel CNN + fixed divergence penalty; 2,514 parameters; test MSE 0.0034996; max divergence 0.21865; RMS divergence 0.03427.",
    "idea": "Scalar-potential CNN + fixed periodic discrete curl; 2,497 parameters; test MSE 0.0037737; max divergence 1.19e-7; RMS divergence 1.76e-8. Identity sanity check max |DCA|=2.22e-16."
  },
  "how_to_run": "/home/maxwelhelp/main/bin/python3 run_experiment.py",
  "files": [
    "run_experiment.py",
    "results.json"
  ],
  "limitations": "This is a one-step supervised toy mapping on a 16x16 periodic grid, not an incompressible PDE rollout or long-horizon energy-drift test. Only one fixed divergence-penalty weight and one seed were evaluated; CUDA convolution failed in the environment and the run fell back to CPU."
}