3D thermal diffusivity reconstruction on PhiFlow synthetic dataset Homogeneous 1,000 samples total
0.96MSE (10^-4)U-NET (FULL)
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| U-NET (FULL)Supervision level=Supervised2026.03 | 0.96 | 24.2 | 94 | 70 | |
| NEFTY (OURS)Model=Full framework2026.03 | 3.66 | 18.48 | 77 | 45 | |
| GRID OPT.Supervision level=Unsupervised2026.03 | 12.61 | 13.99 | 56 | 4 | |
| U-NET (SOUND-ONLY)Supervision level=Supervised2026.03 | 14.73 | 14.83 | 83 | 0 | |
| NeFTY + PE, FA, sigma, HMAblation=PE + FA + sigma + Hard Masking2026.03 | 21.01 | 10.95 | 36 | 24 | |
| NeFTY + PE, FAAblation=PE + Frequency Annealing2026.03 | 29.8 | 8.33 | 16 | 14 | |
| NeFTY + PE, FA, sigmaAblation=PE + FA + standard deviation modeling2026.03 | 31.43 | 9.19 | 26 | 18 | |
| NeFTY + PEAblation=Positional Encoding2026.03 | 126.1 | 4.17 | 12 | 9 | |
| NeFTY BASEAblation=Base model2026.03 | 127.4 | 0.47 | 28 | 3 | |
| PINNSupervision level=Unsupervised2026.03 | 208.3 | -0.24 | 4 | 1 |