Mechanical property prediction of aluminium alloys with varied silicon content using deep learning

This paper presents the development of a deep learning-based model, which can accurately predict the mechanical properties of aluminium alloys with various compositions from microstructural images and clarifies the relationship between the diversity of microstructural images used for model construction and predictive accuracy. Four target mechanical properties—maximum tensile strength, elongation at break, Young's modulus, and 0.2% yield strength—of various aluminium alloys were predicted using a deep learning technique with microstructural images as the input data. Microstructural images were obtained from samples that were cut from ingots produced via gravity die casting or sand mould casting of aluminium alloys with different alloy elements, particularly silicon content levels. In addition, dimensional density, composition information, and X-ray diffraction profiles were utilised as the input data. The models that employed only microstructural images demonstrated high predictive accuracy, with determination coefficients exceeding 0.8 for three mechanical properties. Other input data did not contribute to the enhancement of predictive accuracy. Models that excluded samples with high or low silicon content from the training data could not accurately predict the properties of samples with corresponding silicon content levels. Notably, the silicon content influences the microstructural images, and the absence of similar microstructural images in the training data results in challenges for predicting properties with high accuracy. The incorporation of a more diverse range of microstructural images into the training data suggests the potential to construct a model that can accurately predict the target properties across different microstructures.