Diffractograms are not photos: an ImageNet-only model reads phase fractions from background noise, while physics pre-training locks onto Bragg rings
Synopsis
Synchrotron X-ray diffraction produces a 2D diffractogram per second, and image models seem able to read α and β phase fractions of Ti–6Al–4V from them. Two models reach similar accuracy after fine-tuning on 250 experimental diffractograms, yet robustness tests and attention maps show the ImageNet-only model infers fractions from background and pixel noise, while the model pre-trained on simulated diffractograms relies on Bragg reflections. Accuracy alone does not validate a model.
Interpretation
On α/β phase-fraction regression for Ti–6Al–4V, two networks with similar fine-tuned accuracy rest on entirely different evidence: one on background and pixel noise unrelated to phase, the other on Bragg reflections. The common practice was to treat ImageNet pre-trained weights as the default initialization for scientific imaging tasks and to judge a model by test-set accuracy. Under evaluation that holds out an entire heat-treatment run, the ImageNet-only model diverges when experimental samples are few, while the simulation-trained model stays stable from the first experimental diffractogram.
A curriculum of simulated diffractograms teaches the encoder diffraction-consistent criteria, and a small set of experimental data then bridges the simulation-to-experiment distribution gap. A model trained on simulations without experimental fine-tuning reaches errors comparable to a randomly initialized model trained directly on experimental data, so simulated features do not transfer directly. Across two independent synchrotron 2D diffraction datasets, the simulation-then-fine-tuned model reaches mean absolute errors of 0.045 on one dataset and 0.101 on the other; on 50 randomly selected experimental diffractograms, the zoom and radial-masking differences are statistically significant.
Perspective
The conclusions rest on the specific task of α/β phase-volume-fraction regression in Ti–6Al–4V, with data from two independent synchrotron 2D X-ray diffraction experiments and evaluation that holds out an entire heat-treatment sub-experiment. The authors state that how well the approach works for other systems and tasks depends on their characteristics, and that more realistic simulations may require less experimental fine-tuning.
The simulated diffractograms omit crystallographic texture and its possible correlation with grain size; the authors suggest that preferred-orientation corrections such as March–Dollase could narrow the simulation-to-experiment gap. A model pre-trained by self-supervised masked image modeling on the simulated corpus also shows instability at low experimental sample counts, similar to the ImageNet-only model. The conclusions cover one material system and a two-phase fraction regression task.