HAN-Semi: A Hierarchical Attention Network for Yield Prediction Based on Production Resources in Semiconductor Manufacturing

The rapid advancement of the high-technology industry has intensified the demand for miniaturized semiconductor processes, increasing process complexity and interdependency among process steps. These challenges limit the effectiveness of traditional manual analyses and feature engineering-based machine learning methods, particularly in large-scale production environments with diverse processing conditions. This study addresses these challenges by proposing a hierarchical attention network for yield prediction in semiconductor manufacturing. The proposed method integrates information concerning four critical process conditions (i.e., equipment, chamber, recipe, and mask) across three hierarchical levels (i.e., condition, process, and interaction). This architecture preserves the independent semantics of each condition and captures inter-process dependencies. The final yield prediction is obtained by aggregating outputs from these hierarchical representations. Experiments using real-world data from a Korean semiconductor facility demonstrate that the proposed method outperforms comparative methods, offering superior predictive accuracy and robustness in advanced semiconductor manufacturing processes.