An Explainable Deep Learning Approach for Cycle Time Prediction in Wafer Fabrication

In modern wafer fabrication systems, accurate prediction of job cycle time is vital to production planning reliability, resource allocation, and factory operational efficiency. However, the inherent complexity of manufacturing processes, involving strong job interdependencies and nonlinear, high-dimensional data, limits the effectiveness of traditional prediction methods in practice. To address these challenges, this paper proposes an interpretable deep-learning approach for cycle time prediction that integrates a Transformer neural network with model explainability techniques. First, a dataset is constructed using key production features such as job sizes, bottleneck queue lengths, and work-in-process levels, followed by normalization and time-series preprocessing. Then, a Transformer model is employed to capture the complex relationships among these features, enabling highly accurate cycle time prediction. After model training, both LIME and SHAP methods are applied to reveal the prediction logic from local and global perspectives, thereby identifying the key factors that influence prediction results. Experiments conducted on a real-world wafer fabrication dataset, with comparisons against traditional machine-learning models and deep-learning baseline models, demonstrate the superior performance of the proposed approach in both prediction accuracy and model interpretability. Results also indicate that the proposed approach not only achieves excellent prediction performance on complex production data but also provides technical support for the intelligent and transparent scheduling in wafer fabrication systems, with strong application potential.