Neural network modeling of titanium alloy composition-microstructure-property relationships based on multimodal data

The composition and microstructure of metallic materials determine their mechanical properties. Establishing quantitative “composition-microstructure-property” relationships is essential for material design and optimization. However, materials data are complex and heterogeneous, often containing multi-modal information such as “point-image-curve”. Traditional mathematical models and simulations can no longer meet the needs of advanced applications. Big data technology provides a new approach to materials research, and machine learning tools can offer high-dimensional information in a low-order form. In this study, a neural network-based titanium alloy microstructure identification model and a performance prediction model have been designed. A model for heterogeneous data of titanium alloy composition (data), microstructure (image), and performance (curve) has been established through the fusion of the two models. The cross-dimensional prediction of the material stress-strain curve using composition and microstructure image data has been successfully achieved. A four-step workflow of dataset preparation, pre-processing, model building, and fusion is determined. The task of identifying 1215 images of 14 titanium alloys and predicting 604 performance curves has been completed. This work demonstrates the process of establishing the “composition-microstructure-property” relationship of materials by big data technology.