NeFTY Synthetic Thermography Benchmark

Data and pretrained baseline for Neural Field Thermal Tomography: A Differentiable Physics Framework for Non-Destructive Evaluation (NeurIPS 2026).

Project page | Paper (arXiv 2603.11045) | Code

The benchmark poses label-free 3D inverse heat conduction: given the transient top-surface temperature of a slab after a pulsed heat input, recover the full 3D thermal diffusivity field, including subsurface defects that are never observed directly. Every sample pairs the observed surface thermogram with its exact 3D ground-truth diffusivity, so volumetric reconstructions can be scored voxel by voxel.

Files

File	Content	Size (bytes)	SHA-256
synthetic_1000/dataset.pt	1,000-sample benchmark (homogeneous and layered, with defects)	3,833,858,333	f3c7c73d1b8e61f0dd0bf380cd7b6ecedbf631499ba1ed5de461ecdc20e37f85
synthetic_1000_nodefect/dataset.pt	1,000 defect-free samples (training set of the Sound-Only U-Net baseline)	3,833,858,397	fe504f8374da03f0f48750b6a8bacb720e8abdebef7b77efda62a4de19a3d887
checkpoints/unet_sound_only.tar	pretrained Sound-Only 3D U-Net	84,089,427	11d53863253d1fa75955cdd759b4bee62b8162fd4b1314cfc73513e8833aecff

Benchmark generation

Ground truth is simulated with PhiFlow (Holl and Thuerey, ICML 2024), a finite-volume engine independent of the implicit-Euler solver used by NeFTY, so reconstructions are not evaluated against data produced by their own discretization. PhiFlow integrates the heat equation explicitly with adaptive sub-stepping.

• Domain and grid. Unitless slab [0, 10]² × [0, 1] on a 64 × 64 × 16 grid.
• Boundary conditions. Periodic on the lateral faces and adiabatic on the top and bottom faces.
• Initial condition. A Gaussian post-flash temperature profile localized near the front face, identical for all samples.
• Time axis. 100 recorded frames at Δt = 0.05. The stored surface sequences also include the initial frame, so each holds 101 frames.
• Materials. Bulk diffusivity α_base ~ U(0.1, 0.2) and defect diffusivity α_defect ~ U(0.005, 0.015).
• Defects. One to four ellipsoidal, cylindrical, or box-shaped subsurface defects per sample.
• Bulk configurations. Homogeneous (constant α_base) or layered (three or four strata along z, each with its own bulk diffusivity).

Index layout

synthetic_1000/dataset.pt:

Configuration	Indices	Samples
Homogeneous, 1 defect	0-199	200
Homogeneous, 2 defects	200-399	200
Homogeneous, 3 defects	400-600	201
Homogeneous, 4 defects	601-799	199
3 layers, 1 / 2 / 3 / 4 defects	800-825 / 826-850 / 851-875 / 876-900	101
4 layers, 1 / 2 / 3 / 4 defects	901-924 / 925-950 / 951-975 / 976-999	99

The generator assigns categories from the fractional index i / 1000 in floating point, so some bucket edges fall one sample away from round numbers. Indices 0-799 form the homogeneous split (800 samples) and 800-999 the layered split (200 samples). synthetic_1000_nodefect/dataset.pt uses the same bulk layout without defects: 0-799 homogeneous, 800-900 three layers, 901-999 four layers.

Data format

Each dataset.pt is a Python dictionary saved with torch.save (N = 1000):

Key	Shape	dtype	Content
k_gt	(N, 64, 64, 16)	float32	ground-truth diffusivity on the (x, y, z) cell-center grid
t0_gt	(N, 64, 64, 16)	float32	initial temperature (Gaussian post-flash profile)
surface_gt	(N, 101, 64, 64)	float32	observed top-surface temperature, frames 0 to 100 (equal to surface_gt_phiflow)
surface_gt_phiflow	(N, 101, 64, 64)	float32	PhiFlow ground-truth surface temperature
surface_gt_ours	(N, 101, 64, 64)	float32	NeFTY solver's forward prediction at k_gt (reference only)
resolution	dict		{'x': 64, 'y': 64, 'z': 16}

Usage

import torch
from huggingface_hub import hf_hub_download

path = hf_hub_download(repo_id="n3il666/NeFTY", filename="synthetic_1000/dataset.pt", repo_type="dataset")
data = torch.load(path)

alpha = data["k_gt"][0]          # (64, 64, 16) ground-truth diffusivity of sample 0
surface = data["surface_gt"][0]  # (101, 64, 64) observed surface temperature of sample 0

The code repository downloads every file to the expected location with python scripts/download_data.py, reconstructs individual specimens or the full benchmark with NeFTY and the baselines, and regenerates both datasets with python -m dataset.synthetic and python -m dataset.generate_defect_free (seed 42).

Sound-Only U-Net checkpoint

checkpoints/unet_sound_only.tar holds a 3D U-Net (four-level encoder-decoder with channel sizes 32, 64, 128, 256) that maps the surface temperature sequence to a 3D diffusivity field, with a sigmoid output scaled to [0.003, 0.25]. It is trained with MSE supervision only on the defect-free set, so it serves as a data-driven, zero-shot reference that probes generalization to unseen defects. The file is a dictionary with model_state_dict, optimizer_state_dict, scheduler_state_dict, global_step, and args. Evaluate it on the benchmark with evaluate_unet.py or evaluate_unet_batch.py from the code repository.

Citation

@inproceedings{zhong2026nefty,
  title     = {Neural Field Thermal Tomography: A Differentiable Physics Framework for Non-Destructive Evaluation},
  author    = {Zhong, Tao and Hu, Yixun and Zheng, Dongzhe and Sood, Aditya and Allen-Blanchette, Christine},
  booktitle = {Advances in Neural Information Processing Systems (NeurIPS)},
  year      = {2026}
}

License

Released under the MIT License.

Contact

Tao Zhong (tzhong@princeton.edu), Princeton University.