Accurate reconstruction and prediction of T55511 titanium alloy microstructure using DDPM model and quantitative evaluation

Abstract

The exploration of the relationship between microstructure and properties is essential in materials research. However, traditional methods are often costly and time-consuming, and manual quantitative analysis is subject to high levels of subjectivity, which hinders accurate reflection of microstructural features. In this study, based on a multimodal database of the Solution-Treatment-Aging (STA) process for titanium alloys, we developed a deep learning model for microstructure-property relationships in T55511 alloy by employing a Denoising Diffusion Probabilistic Model (DDPM) with a conditional U-Net as the denoising network. The model generates high-fidelity microstructure images through progressive denoising and adaptively predicts the complex relationships between microstructure and properties based on processing conditions. Results demonstrate that the model produces images with high clarity and strong resemblance to real microstructures. Quantitative information obtained via semantic segmentation shows that the model achieves a mean absolute error (MAE) of ≤30 ​nm for primary phase (αp) and ≤20 ​nm for secondary phase (αs), closely matching real-world data. This study validates the reliability of the DDPM model in generating alloy microstructures and highlights its potential for broad applications, including other alloy systems, process optimization, and materials design.