Diffusion Model for Inverse Design of 7xxx-Series Aluminum Alloys with Desired Property

• Published in: Metals and materials international 2024, 30(7), , pp.1817-1830
• Main Authors: Kim, Sujeong, Zhang, Libin, Kim, Su-Hyeon, Choi, Yoon Suk
• Format: Article
• Language: English
• Subjects: Algorithms; Alloying elements; Alloys; Aluminum alloys; Aluminum base alloys; Characterization and Evaluation of Materials; Chemistry and Materials Science; Clustering; Composition; Datasets; Embedding; Engineering Thermodynamics; Filtration; Heat and Mass Transfer; Inverse design; Machine learning; Machines; Magnesium; Magnetic Materials; Magnetism; Manufacturing; Materials Science; Metallic Materials; Processes; Solid Mechanics; Spatial distribution; Ultimate tensile strength; Zinc; Zirconium; 재료공학
• ISSN: 1598-9623; 2005-4149
• DOI: 10.1007/s12540-023-01610-8

An attempt was made, for the first time, to develop a diffusion model-based inverse design protocol to search composition pools of 7xxx-series aluminum alloys for desired ultimate tensile strength (UTS). Data for Al–Zn–Mg–Cu based 7xxx-series aluminum alloys with minor elements of Mn, Zr, Cr, Ti, Si and Fe, and their UTS were collected and used for training of major machine learning algorithms. Machine learning with composition-UTS dataset showed better predictability (R 2  = 0.82), compared to machine learning with composition-driven elemental property-UTS dataset (R 2  = 0.79). Ensemble model with combination of the two datasets-based models exhibited improved predictability (R 2  = 0.83). Synthetic composition-UTS sets that augmented original composition-UTS dataset were generated through noising and denoising processes of diffusion model, followed by duplication filtering, compositional validity filtering, and large error filtering between ensemble model-predicted UTS and diffusion model-generated UTS. The spatial distribution of original compositions and their UTS were found to be highly correlated with diffusion model-generated synthetic compositions and their UTS in a reduced dimension of t-distributed stochastic neighbor embedding. Visualization of synthetic compositions using the t-distributed stochastic neighbor embedding map allowed the quick and comprehensive assessment of the local clustering behavior of UTS in order to locate particular composition pools for desired UTS. Alloy pools expected to have UTS higher than 750 MPa were identified into three composition groups. These alloy groups revealed relatively high content of Zn and Mg, and were designed so as to obtain extra-enhancement of UTS by adding Mn or Cr. Graphical Abstract