<p>Shrinkage porosity is a common solidification defect in large steel ingots, deteriorating mechanical properties and yield. In this study, a combined framework of numerical simulation and deep learning was proposed to investigate the shrinkage porosity evolution in a 5.5-ton 20Cr13 steel ingot. Solidification sequences of temperature and solid fraction were obtained by numerical simulations, and the Niyama criterion was adopted as the shrinkage indicator. Based on the constructed time-series dataset, a Transformer regression model and an LSTM-Attention regression model were developed to predict shrinkage porosity. Both models achieved accurate predictions, demonstrating the feasibility of applying sequence-based deep learning to metallurgical defect modeling. Furthermore, interpretability analysis using attention weights revealed that solidification stages with solid fraction 0.7–0.95 made the most significant contributions to porosity formation, consistent with the physical mechanism of insufficient feeding in the central mushy zone. Compared with LSTM Attention, the Transformer model exhibited stronger sensitivity to critical solidification stages, highlighting its superiority in capturing defect evolution. The proposed framework provides not only a novel approach for shrinkage porosity prediction but also valuable insights into defect formation mechanisms during ingot solidification.</p>