Abstract: In modern wafer-fabrication systems, accurate prediction of job lead times is critically important for enhancing the reliability of production planning, optimizing resource allocation, and improving overall factory operational efficiency. However, the highly complex nature of the manufacturing process—characterized by strong interdependencies among jobs and by feature data that are both nonlinear and high-dimensional—poses significant challenges to traditional prediction methods, which often encounter performance bottlenecks in practical applications. To address these challenges, this paper proposes an interpretable deep-learning approach for lead-time prediction that integrates a Transformer neural network with model explainability techniques. First, a dataset is constructed using key production features—such as job size, queue length upstream of bottlenecks, and factory work-in-process—and subjected to normalization and time-series preprocessing. Next, a Transformer model is employed to capture and model the complex relationships among these features, enabling highly accurate lead-time predictions. Upon completion of model training, both LIME and SHAP methods are applied to elucidate the model's decision-making process from local and global perspectives, thereby identifying the principal factors influencing prediction outcomes. Experiments conducted on a real-world wafer-manufacturing dataset compare the proposed method against conventional machine-learning models and deep-learning baselines. The results demonstrate that the proposed approach not only achieves a marked improvement in numerical prediction accuracy but also delivers enhanced interpretability. This dual advantage provides robust technical support for the intelligent and transparent scheduling of wafer-fabrication systems and underscores the method's strong potential for practical deployment.