(375p) Quantitative Evaluation of Catalyst Shape in SEM Images Using Gan and Semi-Supervised Semantic Segmentation

The incorporation of Artificial Intelligence (AI) in the field of catalysis engineering is experiencing substantial growth, especially in the analysis of unstructured data, such as SEM images. Scanning Electron Microscopy (SEM) images, as a pivotal source of unstructured data in catalysis engineering, offer insights into catalyst shape, particle size, distribution, and surface texture. The shape of catalysts, closely associated with their catalytic characteristics, can be adjusted through calcination processes, affecting surface facets directly [1]. Traditional methodologies have relied on manual inspection and qualitative assessment of SEM images by catalyst engineers, a process that is not only time-consuming and labor-intensive but also suffers from low accuracy. This research leverages Artificial Intelligence (AI) technology to automate the classification of catalyst particles in SEM images into six distinct shapes—cube, tetrahedron, octahedron, dodecahedron, icosahedron, and sphere—and quantitatively analyzes the distribution of these particle shapes [2]. Consequently, this allows for the quantitative analysis of the distribution of particle shapes, thus addressing the constraints of qualitative evaluation methods by enabling a more efficient and accurate assessment of SEM images, which significantly reducing time and labor costs.

The application of AI to unstructured data research faces challenges, particularly in the early stages due to the high costs associated with acquiring large-scale training datasets. In fields that demand specialized knowledge, such as chemical engineering, access to high-quality image data is exceedingly limited. Furthermore, the quantity of SEM images highlighting shape is particularly scarce, posing a significant constraint on dataset construction. To address these challenges, our research employs Generative Adversarial Networks (GAN) to generate high-quality SEM image data from a limited number of original SEM images [3]. The training data for the GAN model consists of actual images obtained through web crawling on Google, figure images extracted from catalysis literature, and synthetic images created using 3D rendering techniques. Various preprocessing techniques were implemented, such as inpainting to restore text or annotations by leveraging surrounding pixel information, mirror extrapolatation for image size adjustment through symmetrical extension, and augmentation to increase dataset volume, were applied [4]. The prepared data were trained on a DCGAN (Deep Convolutional Generative Adversarial Network) model, employing feature matching and minibatch discrimination techniques, to generate an extensive dataset [5]. Subsequently, the quality of the generated dataset was validated using the Frechet Inception Distance (FID) score.

The dataset generated by the GAN model comprises unlabeled data, while the synthetic dataset created using 3D rendering technology is automatically labeled during its creation process, resulting in a dataset partially consisting of unlabeled and labeled data. This data configuration is suitable for semi-supervised learning approaches. In this research, semi-supervised semantic segmentation techniques were applied, utilizing a method that integrates the U-Net with the Mean Teacher method. This approach segmented the catalyst particles in Scanning Electron Microscopy (SEM) images into six distinct 3D shapes: cube, tetrahedron, octahedron, dodecahedron, icosahedron, and sphere. The accuracy of this segmentation was assessed based on pixel accuracy.

This research presents a shape for overcoming the difficulties associated with acquiring unstructured data and quantitatively representing the distribution of particle shapes in SEM images through the integration of GAN and semi-supervised semantic segmentation within the field of catalysis engineering. By enabling rapid and precise segmentation of catalyst shape, this approach is expected to significantly reduce the cost and time associated with researching the properties related to catalyst shapes, thereby maximizing efficiency.

References

1. Sun, W., X. Li, C. Sun, Z. Huang, H. Xu, and W. Shen, “Insights into the pyrolysis processes of Ce-MOFs for preparing highly active catalysts of toluene combustion,” Catalysts, 9 (8), (2019).
2. Muñoz-Mármol, M., J. Crespo, M.J. Fritts, and V. Maojo, “Towards the taxonomic categorization and recognition of nanoparticle shapes,” Nanomedicine: Nanotechnology, Biology, and Medicine, 11 (2), pp. 457–465 (2015).
3. Creswell, A., T. White, V. Dumoulin, K. Arulkumaran, B. Sengupta, and A.A. Bharath, “Generative Adversarial Networks: An Overview,” IEEE Signal Processing Magazine 35, 2018, 53–65.
4. Navab, N., J. Hornegger, W.M. Wells, and A.F. Frangi, “LNCS 9351 - Medical Image Computing and Computer-Assisted Intervention – MICCAI 2015,” (2015).
5. Salimans, T., I. Goodfellow, W. Zaremba, V. Cheung, A. Radford, and X. Chen, “Improved Techniques for Training GANs,” (2016).
6. Chen, X., Y. Yuan, G. Zeng, and J. Wang, “Semi-Supervised Semantic Segmentation with Cross Pseudo Supervision,” (2021).