TiM3: Titanium Alloy Microstructure Generation Following Multiple Mechanical Properties

Zhixin Zhang, Enhua Guo, Jinzhe Liu, Zhang zhixuan, Fusheng Jin, Fan Qunbo

Keywords: image generation, ai for science, titanium alloys, micro structure

Abstract: Generating titanium alloy microstructure images according to the required mechanical properties can guide the optimization of materials and save lots of resources that would otherwise be consumed by traditional try-and-test material experiments. Existing microstructure image generation methods mainly focus on unconditional or discrete conditional applications, failing to support multiple continuous mechanical properties, such as Rockwell Hardness. Moreover, a crucial factor that determines the properties of titanium alloys, metallography, is still underexplored, which leads to the risk of unrealistic distorted structures. In this work, we propose **TiM3**, a two-stage coarse-to-fine latent diffusion model for **Ti**tanium alloy **M**icrostructure generation following **M**ultiple **M**echanical properties. We embed required mechanical properties using a probabilistic encoder to raise model robustness to unseen conditions and adopt cross-attention to guide the generation process. To enable the model to focus on the metallography of titanium alloys, we separate the generation into two stages. The first stage produces coarse metalloraphic structures as intermediate representations, and the second stage complements these with fine-grained details. To evaluate models' continuous generation capability in the high-dimensional mechanical property space, we design a property sampling algorithm to balance generalizability testing and property authenticity. TiM3 shows outstanding microstructure image quality, diversity, and property accuracy in both quantitative and qualitative experiments.

Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)

Submission Number: 7727