Self-supervised denoising lets an unsupervised pipeline map crystals in noisy 4D-STEM data

This method does without clean training images ^[5], and those images are hard to get for the materials the authors single out. Beam-sensitive specimens such as polymers give low signal-to-noise diffraction data, and the abstract calls their analysis particularly challenging ^[3]. A denoiser that learns from the noisy data alone suits that situation.

4D-STEM records a diffraction pattern at every probe position across a region of interest ^[1]. To work on that kind of data, the denoiser borrows from the patterns recorded at neighbouring positions ^[6]. This makes sensible use of the data's structure. It also assumes neighbours are alike. I'd expect the method to do best inside large grains and to face its hardest test at boundaries and small features, where adjacent patterns differ.

The second stage sorts patterns without labels. It is designed to be invariant to in-plane rotation of the crystal ^[7]. Without that property, an unsupervised method could split one crystal component into several groups just because it appears at different angles. The authors say the invariance is what lets the pipeline identify components effectively ^[7].

The evaluation uses two controls of different strength. The synthetic data, tested across a range of noise levels ^[8][9], comes with a known right answer. The experimental data was checked against an expert analysis from a previous study ^[10]. Agreement there shows the pipeline can reproduce a trained human's reading. The minor components it found that the expert analysis had missed ^[11] are harder to score. With only one reference, a new feature could be a real phase or something the denoising step produced. The abstract does not give error rates or name the experimental specimen. It does not quantify the cost saving the authors call significant ^[12], or say how the new components were checked against denoising artefacts.

In my view the self-supervised denoiser is the part most likely to matter to people imaging polymers. That holds only if the full paper shows it working at the lowest doses those samples tolerate. The paper, by Kamijo, Shiga, Kanomi and colleagues, appears in npj Computational Materials ^[13]. It was funded by JST BOOST, JST CREST and JSPS KAKENHI grants, and the authors declare no competing interests ^[14]. It is open access under a CC BY 4.0 licence ^[15].