EUVIMEDEUVIMED entdecken
Crossref

An end-to-end glass-forming ability prediction model based on the Mamba architecture

Junjie Zeng, Yongchao Liang, Gongmin Wei, Shengli He, Can Wu, Jiamu Zhao, Qian Chen

Physica Scripta · 20.03.2026 · 101 (11) · 116001

DOI: 10.1088/1402-4896/ae4ffc ↗

Kurz gefasst

Fachartikel in der Zeitschrift Physica Scripta, online erschienen am 19.03.2026. Verfasst von Junjie Zeng und 6 weiteren Autor:innen. Die Arbeit nennt 49 Literaturangaben. Gefördert von: Guiyang City Science and Technology Plan Project, Natural Science Special (Special-term Professor) Research Foundation of Guizhou University, National Natural Science Foundation of China.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

In recent years, machine learning (ML) approaches that predict glass-forming ability (GFA) from characteristic temperatures have opened new avenues for the development of metallic glasses (MGs). However, these temperatures are available only after sample preparation, which is time-consuming and resource-intensive. To overcome this limitation, we encode alloy compositions to jointly represent element identity and fraction and, on this representation, adopt a sequence-modeling paradigm to capture long-range dependencies and global inter-element relations. Nevertheless, conventional sequence models lack input-conditioned adaptivity and struggle with the complexity and heterogeneity of multicomponent alloys. Accordingly, we propose the Gated-channel Attention Mamba (GA-Mamba), a hybrid architecture that embeds a MambaBlock to adapt parameters to the input and integrates a composition-oriented Gated-channel Attention (GA) module to align and reweight elemental embeddings, thereby strengthening representation and prediction for complex alloy systems. Compared with baselines, GA-Mamba more effectively captures nonlinear inter-element interactions and improves prediction accuracy and generalization. To mitigate data scarcity and distribution shift, we employ three oversampling strategies, including Gaussian noise (GN), Synthetic Minority Over-sampling Technique (SMOTE), and Mixup, to better balance the training data; Mixup yields the largest gain, and the model attains R 2 = 0.8722 on the test set. Predictions for the Cu-Ca-Mg and Cu-Al-Zr ternary systems further demonstrate strong generalization.

Förderung

  • Guiyang City Science and Technology Plan Project
  • Natural Science Special (Special-term Professor) Research Foundation of Guizhou University
  • National Natural Science Foundation of China
  • ,Innovation and project of Guizhou Entrepreneurship education University
  • Project of Teaching Content and Curriculum System Reform in Higher Education Institutions of Guizhou Province
  • Guizhou Provincial Department of Education Youth Science and Technology Elite Talent Project

Zitieren

APA

Zeng, J., Liang, Y., Wei, G., He, S., Wu, C., Zhao, J., & Chen, Q. (2026). An end-to-end glass-forming ability prediction model based on the Mamba architecture. Physica Scripta, 101(11), 116001. https://doi.org/10.1088/1402-4896/ae4ffc

Vancouver

Zeng J, Liang Y, Wei G, He S, Wu C, Zhao J, et al. An end-to-end glass-forming ability prediction model based on the Mamba architecture. Physica Scripta. 2026;101(11):116001. doi:10.1088/1402-4896/ae4ffc

BibTeX
@article{Zeng2026,
  author = {Zeng, Junjie and Liang, Yongchao and Wei, Gongmin and He, Shengli and Wu, Can and Zhao, Jiamu and Chen, Qian},
  title = {An end-to-end glass-forming ability prediction model based on the Mamba architecture},
  journal = {Physica Scripta},
  year = {2026},
  volume = {101},
  number = {11},
  pages = {116001},
  doi = {10.1088/1402-4896/ae4ffc}
}
Kostenlos starten →