EUVIMEDEUVIMED entdecken
Crossref

A Decoder-Only Transformer Model for Robust Bragg Peak Detection in Noisy Liquid-Cell 4D-STEM Data

Zhijing Zhang, Kwanghwi Je, Erik Thiede

Microscopy and Microanalysis · 27.07.2026 · 32 (Supplement_1)

DOI: 10.1093/mam/ozag053.043 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Microscopy and Microanalysis, online erschienen am 27.07.2026. Verfasst von Zhijing Zhang und 2 weiteren Autor:innen. Die Arbeit nennt 5 Literaturangaben.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Operando electrochemical liquid-cell scanning transmission electron microscopy (EC-STEM) is a powerful technique for monitoring nanoscale structural evolution in real time during electrochemical reactions [1]. Four-dimensional scanning transmission electron microscopy (4D-STEM) records a diffraction pattern at every probe position, enabling quantitative phase identification and orientation mapping to extract crystallographic information [2]. EC-STEM can be combined with 4D-STEM, which forms a compelling operando platform for uncovering fundamental mechanisms that govern the performance of energy materials [3]. In practice, however, robust quantitative analysis in liquid-cell 4D-STEM is often limited by a simple step: reliable Bragg peaks detection. Under low-dose conditions in the liquid environment, true Bragg disks can be weak and sparse. Moreover, the liquid environment can generate numerous false positives, leading to large errors in downstream data processing. To address this limitation, we present a decoder-only transformer deep learning model that reframes Bragg disk detection as a sequence modeling problem governed by physical rules. We first generate an initial pool of peak candidates using a loose intensity-based threshold [4]. This ensures a high recall of potential Bragg disks, albeit with a significant number of false positives derived from the noisy background. These candidates are then tokenized into a sequence of coordinates parameterized by radial position (qr), angular position (qθ), and intensity (I), which naturally aligns with the physics of diffraction, where qr corresponds to lattice spacing (1/d) and qθ to lattice orientation. Candidates are ordered by descending intensity and passed through the decoder-only transformer architecture, which processes the candidates via a causal masking mechanism [5]. The causal masking mechanism delivered by the decoder enables the model to perform progressive geometric refinement. By processing peaks in order of decreasing intensity, the model effectively anchors its crystallographic hypothesis on the brightest, highest-confidence features first, which is suitable for the noisy liquid-cell 4D-STEM dataset. When the model evaluates a subsequent, lower-intensity candidate, the masked self-attention mechanism forces it to assess that candidate’s geometric compatibility, such as crystal symmetry, against the stronger peaks observed previously. This allows the network to differentiate between a weak true peak (which forms a coherent lattice with the high-intensity anchors) and a random noise artifact (which appears geometrically isolated). The model assigns a probability score to each token based on this cumulative structural context, enabling the precise filtration of the initial noisy candidate list. We validated this approach using a library of simulated diffraction patterns of Cu fcc crystal integrated with experimental background noise (μbg = 1 to 40). The decoder-only transformer trains stably across multiple background conditions and can learn to suppress background-driven false positives without sacrificing recall of weak true peaks. In our initial experiments, training and validation loss curves decrease smoothly for several background-scaling settings, supporting the feasibility of the causal-attention design under heavy noise. Qualitatively, the model’s filtered peak sets preserve reciprocal-lattice structure more consistently than traditional algorithms like template matching when background artifacts dominate, which is expected to improve downstream tasks such as orientation mapping and phase identification. On high-background test data, our model maintains strong recall while filtering out the vast majority of false positives. This confirms that the model effectively learns the underlying physical rules of crystallography. By using these rules to filter out noise, it serves as a reliable foundation for accurate orientation mapping, even in difficult, low-dose imaging conditions in the liquid-cell EC-4D-STEM.

Zitieren

APA

Zhang, Z., Je, K., & Thiede, E. (2026). A Decoder-Only Transformer Model for Robust Bragg Peak Detection in Noisy Liquid-Cell 4D-STEM Data. Microscopy and Microanalysis, 32(Supplement_1). https://doi.org/10.1093/mam/ozag053.043

Vancouver

Zhang Z, Je K, Thiede E. A Decoder-Only Transformer Model for Robust Bragg Peak Detection in Noisy Liquid-Cell 4D-STEM Data. Microscopy and Microanalysis. 2026;32(Supplement_1). doi:10.1093/mam/ozag053.043

BibTeX
@article{Zhang2026,
  author = {Zhang, Zhijing and Je, Kwanghwi and Thiede, Erik},
  title = {A Decoder-Only Transformer Model for Robust Bragg Peak Detection in Noisy Liquid-Cell 4D-STEM Data},
  journal = {Microscopy and Microanalysis},
  year = {2026},
  volume = {32},
  number = {Supplement_1},
  doi = {10.1093/mam/ozag053.043}
}
Kostenlos starten →