Automatic determination of quasicrystalline patterns from microscopy images
Tano Kim Kender, Marco Corrias, Cesare Franchini · 2025
Read Automatic determination of quasicrystalline patterns from microscopy images by Tano Kim Kender, Marco Corrias, Cesare Franchini (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2503-05472.txt.
Introduction
Kender, Corrias, and Franchini presented a method in twenty twenty-five for spotting quasicrystal patterns in microscope images automatically. Quasicrystals are solids with ordered atoms that never repeat in a regular grid. Standard software for crystal lattices cannot read their arrangements.
Their method uses image feature recognition plus machine learning, tuned for tilings with eight-, ten-, and twelve-fold symmetry. Clustering finds the atom positions first. Then neighbor analysis and border tracing build the tiling from the links between atoms. A support vector machine cleans up the results, reaching accuracy in line with past reports. The code now ships in the open-source AiSurf package, with a statistical check of its performance included.
Catalog record
| CREATOR | Tano Kim Kender, Marco Corrias, Cesare Franchini |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 25 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2503-05472
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Automatic determination of quasicrystalline patterns from microscopy images Preprint: APS/123-QED Tano Kim Kender, Marco Corrias, Cesare Franchini Abstract Quasicrystals are aperiodically ordered solids that exhibit long-range order without translational periodicity, bridging the gap between crystalline and amorphous materials. Due to their lack of translational periodicity, information on atomic arrangements in quasicrystals cannot be extracted by current crystalline lattice recognition softwares. This work introduces a method to automatically detect quasicrystalline atomic arrangements and tiling using image feature recognition coupled with machine learning, tailored towards quasiperiodic tilings with 8-, 10- and 12-fold rotational symmetry. Atom positions are identified using clustering of feature descriptors. Subsequent nearest-neighbor analysis and border following on the interatomic connections deliver the tiling. Support vector machines further increase the quality of the results, reaching an accuracy consistent with those reported in the literature. A statistical analysis of the results is performed. The code is now part of the open-source package AiSurf. I Introduction Quasicrystals are a class of aperiodic materials showing long-range ordering. While quasicrystalline patterns have appeared in various early artworks and architectural designs Al Ajlouni 2013 , the first experimental observation of a ten-fold electron diffraction pattern, revealing non-periodic tilings of the plane, was reported by Dan Shechtman et al. in 1984 Shechtman et al. 1984 . Since then, quasicrystalline systems have been extensively studied for their natural beauty and unusual characteristics, leading to the discovery of several remarkable physical properties Louzguine-Luzgin and Inoue 2008 . Some quasicrystals have been reported to show high electrical resistivity Kashimoto et al. 2000 , and can be suitable for hydrogen storage Takasaki and Kelton 2006 ; Gibbons et al. 2004 ; the lack of translational symmetry makes the quasicrystals usually hard and brittle at room temperature Edagawa 2001 . Recent studies report bulk superconductivity Takemori 2024 ; Tokumoto et al. 2024 , nematic superconductivity Liu et al. 2024 and asymmetric phonon intensities Matsuura et al. 2024 . It is well known that crystalline lattices are only compatible with 1-, 4- and 6-fold rotational symmetries Hoffmann 2020 , and their ideal structure can be described in terms of a single unit cell. Quasicrystalline structures, however, do not possess the periodicity paradigm of classical crystallography as 5-, 8-, 10- and 12-fold rotational symmetries are allowed Stadnik 2012 : they show Bragg spots in the diffraction pattern, but no translational periodicity Yamamoto 1996 . This poses major challenges for the crystallographical description of such structures, but also attracts the attention of a vast scientific community that tries to identify structure-property correlations Dolinšek 2012 . Not describable using single unit cells, quasiperiodic patterns can be generated by packings, coverings, and tilings Steurer and Deloudi 2008 . Recent studies show promising steps towards a more generalized, formal description of the tilings Imperor-Clerc et al. 2024 ; Impéror-Clerc et al. 2021 . These tilings, such as the renowned Penrose tiling, consist of primitive tiles that can be arranged into an aperiodic pattern by following matching rules Grünbaum and Shephard 1987 . They can be interpreted as the projections of lattices from a higher-dimensional space Yamamoto 1996 . In recent years, the first oxide quasicrystal (OQC) has been observed Förster et al. 2013 ; Förster et al. 2020 and other studies followed on that direction Förster et al. 2016 ; Zollner et al. 2020 ; Wührl et al. 2023 ; Merchan et al. 2022 , hinting for a promising, unexplored area of surface science. Atomically resolved images can reveal several mechanisms and features encompassing crystallographic information, domain orders, molecular adsorption, and distribution of defects. During the last decade, the interest in automated analysis and pattern recognition for microscopy images has significantly grown Kalinin et al. 2023a ; Kalinin et al. 2022 . Several different approaches are proposed in the literature to extract information from these images. WSXM Horcas et al. 2007 and ImageJ Collins 2007 are popular choices for manual analysis. Machine learning has seen an exponential increase in popularity in recent years, and its applications in the field of microscopy are now several Kalinin et al. 2023b ; Kalinin et al. 2023c ; Pregowska et al. , such as convolutional neural networks Patrick et al. 2023 ; Madsen et al. 2018 ; Leitherer et al. 2023 ; Ziletti et al. 2018 and graph neural networks Li et al. 2023 . Unsupervised methods also saw an increase in popularity, starting from clustering algorithms Corrias et al. 2023 , variational autoencoders Ziatdinov et al. 2023 ; Biswas et al. 2023 , support vector machines Guo et al. 2021 , and others Somnath et al. 2018 . Machine learning and data-driven approaches have been applied to identify, predict, and design quasicrystalline structures Uryu et al. 2024 ; Liu et al. 2021 ; Liu et al. 2023 ; Fujita et al. 2024 . However, to the best of our knowledge, no publicly available methods have been specifically designed to extract quasicrystalline patterns from microscopy images. In this work, we introduce a pattern recognition tool tailored for quasicrystalline surfaces, based on supervised and unsupervised machine learning. Its performance is tested on atomically resolved scanning tunneling microscopy (STM) and High-angle annular dark-field imaging (HAADF) images of polygonal quasicrystal and approximant surfaces. The workflow is presented in the next section, followed by results and discussions. Data, code, and documentation are available as part of the open-source AiSurf Python package Corrias et al. 2023 . II Methods Figure 1: Key steps in the quasiperiodic pattern recognition process on an STM image of a 12-fold rotationally symmetric BaTiO 3 quasicrystal on Pt(111). Adapted with permission. Förster et al. 2013 Copyright 2025, Nature. (a) Detected atomic keypoints shown by green circles with radii indicating the reduced keypoint size. Bottom right: nearest-neighbor distribution with k=7 , color-coded by cluster labels with different distance from the origin. The black arrows represent the symmetry vectors pointing to the 12 clusters of relative nearest-neighbor locations in the dominant distance cluster (blue). (b) Tiling outline created by connecting each atomic keypoint center to the center of all its neighbors in the dominant distance cluster. A border-following algorithm is applied to this image to extract contours. (c) Final tiling after completing missing point search in ”open” contours. Tile colors are assigned based on their internal angles. Holes in the tiling stem from vacancies in the quasicrystal. Pattern recognition utilizes image feature extraction as a means of finding atomic positions. To this end, the Scale Invariant Feature Transform (SIFT) Lowe 2004 algorithm is employed. By repeatedly convolving atomically resolved images with a Gaussian function with increasing \sigma and taking the difference between two consecutive convolutions, a bandpass filter is applied, acting as a feature enhancer. From these stacks of Difference of Gaussian (DoG) images, split by limiting the increase of \sigma to double its initial value and downsampling before continuing the convolutions, extremum points are extracted as the features of the images, referred to as ”keypoints”. Although the SIFT parameters require some tuning for optimal results depending on the image quality and resolution, the default values are set to ensure meaningful initial results. Filtering out keypoints outside of a set range around the median keypoint size discards the first batch of non-atomic keypoints. The remainder, atomic keypoints and non-atomic keypoints of atomic size,