3D thermal diffusivity reconstruction on PhiFlow synthetic dataset Layered Composite
0.0003MSEU-NET (FULL)
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| U-NET (FULL)Supervision level=Supervised2026.03 | 0.0003 | 20.03 | 90 | 68 | |
| NEFTY (OURS)Model=Full framework2026.03 | 0.0009 | 15.88 | 74 | 37 | |
| U-NET (SOUND-ONLY)Supervision level=Supervised2026.03 | 0.001 | 15.42 | 88 | 0 | |
| GRID OPT.Supervision level=Unsupervised2026.03 | 0.0015 | 13.27 | 57 | 3 | |
| NeFTY + PE, FA, sigma, HMAblation=PE + FA + sigma + Hard Masking2026.03 | 0.0023 | 11.26 | 33 | 22 | |
| NeFTY + PE, FA, sigmaAblation=PE + FA + standard deviation modeling2026.03 | 0.0036 | 9.27 | 24 | 14 | |
| NeFTY + PE, FAAblation=PE + Frequency Annealing2026.03 | 0.0044 | 8.81 | 18 | 14 | |
| NeFTY + PEAblation=Positional Encoding2026.03 | 0.0087 | 6.15 | 13 | 9 | |
| PINNSupervision level=Unsupervised2026.03 | 0.02 | 1.42 | 4 | 2 | |
| NeFTY BASEAblation=Base model2026.03 | 0.0222 | 2.89 | 24 | 2 |