Acombined approach of numerical simulation and deep learning was proposed to investigate shrinkage porosity defects during the solidification process of steel ingots. A Transformer neural network model with attention mechanisms was developed for predicting shrinkage porosity. First, finite element numerical simulations were conducted for the solidification process of an 8-ton 9Cr3Mo steel ingot to obtain time series data such as the nodal temperature and solid fraction. The simulation data were subsequently used as input to construct a Transformer regression model with multi-head self-attention mechanisms to predict shrinkage porosity. Finally, the attention weights of the model were analysed to reveal its focus on key features such as the solid fraction at different stages of the solidification process. The results show that the model automatically focuses on the late solidification stage of the steel ingot, which aligns with the physical mechanism of shrinkage formation, providing a data-driven perspective for identifying shrinkage-prone regions.